Beyond word error rate: Multidimensional evaluation of ASR performance for digital speech biomarkers

Beyond word error rate: Multidimensional evaluation of ASR performance for digital speech biomarkers | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Beyond word error rate: Multidimensional evaluation of ASR performance for digital speech biomarkers Thayabaran Kathiresan, Zona Ling, Loretta Gasparini, Stephanie Eliza Piccolo, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8824529/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Automatic speech recognition (ASR) is increasingly used in digital health, yet its reliability for populations with atypical speech, such as people with dementia, is not well characterised beyond word error rate (WER). We evaluated eight ASR systems on speech from adults with dementia and healthy older adults using WER, part-of-speech sequence agreement (POS-sqWER), part-of-speech distribution mismatch (POS-MDE), sentence-embedding cosine distance, and stutter detection error (SDE). Mixed-effects models with Tukey-corrected contrasts were used for comparison. Performance was consistently poorer for dementia speech across all metrics. AWS and CrisperWhisper showed relatively strong lexical, semantic, and syntactic fidelity, whereas Google and Meta exhibited lower accuracy. Other systems showed intermediate performance with rankings varying by metric. POS-MDE revealed syntactic distortions not captured by WER or POS-sqWER. SDE performance was low across systems. Multidimensional evaluation reveals clinically relevant linguistic distortions obscured by WER alone, supporting the need for population- and task-specific ASR validation in digital health research. Health sciences/Biomarkers Biological sciences/Computational biology and bioinformatics Health sciences/Health care Health sciences/Medical research Biological sciences/Neuroscience Automatic speech recognition Natural language processing Speech Language Dementia Digital biomarkers Clinical research Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Automatic Speech Recognition (ASR) technologies continue to advance, achieving near-human transcription accuracy in well-controlled settings 1 , showing robustness in adverse, noisy conditions 2 . These advances have spurred widespread integration of ASR in consumer applications, enterprise solutions, and in healthcare 3 . As digital health ecosystems expand, there is growing interest in leveraging ASR for clinical applications such as remote monitoring, diagnostic support, automated documentation, and telehealth services 4 . Speech is a rich marker of neurological and cognitive health, offering a non-invasive and accessible clinical assay 5 – 7 . However, we do not know how accurately ASR systems process real-world clinical speech, which remains a critical barrier. Unlike the clean and fluent speech used to train most commercial ASR models, clinical speech is often marked by disfluencies, atypical prosody, articulatory breakdowns, and phonatory disturbances. These features are prominent in individuals with neurodegenerative conditions such as Parkinson’s disease 8 Huntington’s disease 9 , multiple sclerosis 10 , Amyotrophic Lateral Sclerosis (ALS) 11 , and various forms of dementia 12 , where speech and language performance varies with disease status. In cases of atypical communication resulting from primary progressive aphasia (PPA) and fronto-temporal dementia, conventional ASR systems may not accurately transcribe words and phrases. PPA and FTD can lead to reduced intelligibility, uncommon phonological patterns, dysfluent or agrammatical phrasing or articulatory imprecision 13 , 14 , but also to handle ungrammatical utterances. Most available models are trained on clean, non-disordered speech and language, so they tend to produce the most statistically plausible sentence rather than faithfully render atypical productions 15 . In this case, rather than transcribing an utterance as it was spoken the model is likely to “correct” the sentence by omitting or misclassifying tokens that may carry clinical relevance 16 – 19 . Failure to transcribe repeated syllables, mispronounced words, or pauses may obscure patterns crucial for detecting disease progression or symptom severity. While recent ASR architectures, particularly large-scale, transformer-based models like OpenAI’s Whisper and Meta’s MMS have demonstrated improved robustness to diverse speech 20 , 21 , their real-world performance on disordered clinical speech remains under-investigated. Evaluating these systems using conventional metrics such as Word Error Rate (WER) provides a partial view of their performance 19 . WER captures surface-level discrepancies but falls short of revealing how well the ASR preserves syntactic and grammatical structure: an aspect critical in clinical Natural Language Processing (NLP) tasks such as information extraction, temporal reasoning, and diagnosis classification. To address these challenges, we conducted a multi-faceted evaluation of eight state-of-the-art commercial ASR systems on a curated dataset comprising speech samples from individuals with dementia and healthy older adults (Fig. 1 ). Our evaluation strategy integrated four complementary analyses. First, we applied Word Error Rate (WER), the standard benchmark for transcription accuracy, which measures token-level discrepancies through insertions, deletions, and substitutions. Second, we implemented Part-of-Speech (POS) distribution analysis, a syntactic-level approach that compares the frequency of grammatical categories such as nouns, verbs, and auxiliaries between ASR outputs and the ground truth. This method highlights whether the structural composition of utterances is retained, even when individual word choices differ. Third, we introduced a novel metric, POS-sqWER (Part-of-Speech sequence Word Error Rate), which blends POS tagging with the sequential alignment logic of WER. By computing the Levenshtein distance (a measure of mismatches) 22 between of POS tags from ASR and reference transcripts, POS-sqWER captures both syntactic fidelity and positional accuracy. Fourth, we implemented a stutter detection analysis to evaluate the ability of ASR models to accurately identify and reflect dysfluent speech patterns. [Insert Fig. 1 here] Finally, we incorporated a sentence-level embedding similarity analysis using large language models to evaluate semantic coherence between ASR outputs and the ground truth. This approach captures higher-level linguistic alignment, offering a metric sensitive to the overall meaning of an utterance even in the presence of lexical or syntactic noise. By comparing vector embeddings of sentences, we assessed the extent to which ASRs preserve potentially clinically relevant performance (e.g. word fragments, repetitions) and the original meaning of the speaker’s utterances, which are crucial for diagnosing or monitoring conditions that affect speech or language. Related Work Automatic speech recognition (ASR) is increasingly used to support clinical documentation and analysis of patient-clinician conversations, yet most evaluations often emphasize headline word error rate (WER), which can treat all errors equally and may not capture differences in clinical impact or meaning preservation 23 . Reviews of clinical ASR report wide WER ranges (~ 18–63%) in realistic settings 24 , 25 , underscoring that a single aggregate error rate obscures what goes wrong and how it affects downstream clinical tasks. To address the limitations of WER, meaning- and task-oriented evaluations have emerged. “Semantic-WER” and related semantic distance metrics capture transcript usefulness by rewarding meaning preservation over literal n-gram overlap 26 . In parallel, studies pairing ASR with medical NLP that evaluate clinical entity-level accuracy (e.g., medications, problems) have shown that two systems with similar WER can differ substantially in clinical utility, especially on accented speech and specialized terminology 27 – 29 . Despite these advances, prior clinical ASR evaluations rarely quantify how transcripts preserve the grammatical scaffolding clinicians rely on (e.g., auxiliaries, modals, negation cues). Parts-of-speech in ASR research often attempts to describe error classes, not to score ASR output itself. For example, POS-aware analyses have been used to study the impact of different word classes on conversational user-modelling, but they stop short of defining POS-centric evaluation measures for ASR quality 30 . We did not find clinical ASR studies that (i) measure distributional shifts in POS categories or (ii) penalize local POS-sequence disruptions as first-class metrics. This paper fills that gap by examining the accuracy of contemporary ASR systems in clinical speech cohorts by introducing two complementary metrics: POS-MDE (mean distribution error over POS categories) quantifies whether an ASR preserves the global grammatical profile of clinical discourse, and POS-sqWER (sequence-aware WER on POS tags) penalizes local syntactic mis-orderings that can flip meaning. Used alongside WER and entity-centric scores, these metrics provide a clinically relevant lens for selecting ASR systems that not only recognize words but also preserve grammatical signals crucial for interpretation, temporality, and recognition of clinically relevant communication features. Methods Participants and speech tasks All language samples were conducted in English, and most participants were native speakers of Australian English. Participants completed a variety of structured and semi-structured speech elicitation tasks, which differed between groups depending on the dataset source. Participants described their daily activities or personal experiences (Monologue) and completed a picture description task (The Cookie Theft Description). Detailed demographic information is provided in Table 1 . All audios were recorded in clinical or research environments at 44.1 kHz sampling rate and underwent manual quality control for completeness. [Insert Table 1 here] Table 1 Participants demographic and experiment tasks. This study was approved by the Eastern Health Human Research Ethics Committee (HREC), Reference No. E21-1314-23343 (ERM ID: 23343), and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants or their legally authorised representatives prior to participation. Automatic speech recognition systems To evaluate transcription performance across various automatic speech recognition (ASR) systems, we employed eight contemporary ASR models, representing both open-source and proprietary systems (shown in Table 2 ). These systems were selected based on their recent performance in general-purpose and clinical domains, multilingual capabilities, and model architecture diversity: [Insert Table 2 here] Table 2 Description of ASR models evaluated in this study. ASR Model Description AWS Transcribe 31 Amazon’s cloud-based ASR platform, evaluated in its default medical transcription settings. OpenAI Whisper (large) 20 Transformer-based multilingual model trained on 680,000 hours of audio; robust to noise and accents. NVIDIA Nemo 32 Research-grade transformer ASR model with general-purpose transcription capabilities. AssemblyAI 33 Commercial ASR API with support for medical domain customization and disfluency filtering. Microsoft Phi-4 34 Instruction-following multimodal ASR model capable of speech-to-text with advanced contextual understanding. CrisperWhisper 35 , 36 Fine-tuned Whisper variant optimized for clinical and disfluent speech, designed for medical applications. Meta MMS-1B-all 21 , 37 Multilingual ASR model trained on over 1,000 languages, developed by Meta (Facebook). Google ASR 38 Google’s medical transcription model using the medical conversation setting for healthcare dialogue. We used the raw ASR model transcripts in our analysis, with no modifications or corrections, to allow for accurate comparison with the manual ground truth transcriptions. Manual Transcription Annotation The transcription framework was based on guidelines by 39–41 . The framework was designed to be compatible with Universal dependencies (UD) and TalkBank 42 , who each provide tools for language sample analysis. Our framework guides users on how to represent disfluencies. For example, word fragments are represented with a hyphen (‘I w- went to the park’), and incomplete utterances are following by an ellipsis (‘why don’t you put it … don’t do that!’). These serve as critical ground truth markers for evaluating disfluency detection performance in our subsequent analyses. The manual ground truth transcriptions were produced and quality-checked by three speech pathologists with further quality checks by a clinical linguist. For a random sample of 49 of the transcripts, two transcribers independently produced a transcript by listening to the audio recording and following the transcription framework. They discussed with each other any discrepancies between their versions and resolved them through consensus to generate a finalized, accurate transcript. A further 49 were then transcribed by one transcriber by listening to the audio recording and following the transcription framework. Initial transcripts for all audio files had previously been generated by crowd-sourced workers via Amazon Mechanical Turk (MTurk); these had later been reviewed and corrected by in-house annotators. These had not followed any specific framework and typically did not include clinically relevant disfluencies like word fragments or repetitions, so they were not fit for purpose. The transcribers corrected the remaining 210 texts initially formulated by MTurk transcription by listening to the audio recording and following our transcription framework. This was designed to reduce transcription time, while maintaining high quality. Preprocessing and Transcript Normalization All ASR-generated and manually transcribed texts were subjected to a standardized set of normalization and pre-processing procedures to ensure consistency and comparability across evaluation metrics. First, all text was converted to lowercase to mitigate case-based mismatches. Punctuation and special characters such as commas, periods, brackets, and other non-linguistic symbols were removed to prevent unintended lexical artifacts. Common filler words, including vocalized hesitations like “um” and “uh,” were retained to preserve disorder-specific speech characteristics in the transcription. Any extraneous whitespace introduced during transcription or formatting was trimmed. Numeric values were converted to their corresponding textual forms (e.g., “2” to “two”) to maintain alignment consistency across tools and methods. Finally, all transcripts were sorted by filenames in ascending alphanumeric to ensure consistent matching between manual and ASR transcriptions throughout the analysis pipeline. Evaluation Metrics To evaluate and compare ASR performance against the manually transcribed ground truth, we employed a combination of surface-level and linguistic-level metrics, listed below and each explained further in the following sections. Word Error Rate (WER) Word Error Rate (WER) is a widely accepted and standardized metric 43 used to evaluate the overall transcription accuracy of Automatic Speech Recognition (ASR) systems. It computes the minimum number of insertions (I), deletions (D), and substitutions (S) required to transform the ASR-generated text into the ground truth, normalized by the total number of words in the ground truth (N). WER offers a high-level, interpretable measure of ASR performance, enabling straightforward comparisons across systems. It is easy to calculate and widely reported, making it ideal for benchmarking tasks. WER provides a quick estimate of transcription fidelity. However, it does not account for semantic or syntactic nuances. It treats all errors equally without differentiating between significant and negligible changes (e.g., omitting a repeated word vs. a replacing a noun with an unrelated verb that changes both the grammatical structure and meaning of the sentence). Moreover, WER operates at the word level and fails to capture subtler structural properties of language or sentence coherence, which are particularly crucial for clinical samples. Part-of-Speech (POS) Mean Distribution Error (POS-MDE) To assess the syntactic fidelity of automatic speech recognition (ASR) outputs, we conducted a Part-of-Speech (POS) distribution analysis, which evaluates how closely the grammatical structure of ASR transcriptions aligns with that of manually transcribed ground truth texts. This analysis focuses on the frequency distribution of 13 POS categories, including nouns, verbs, and conjunctions (see Supplementary Table 1). Each transcript was POS-tagged using a natural language processing library Stanza 44 , and the proportion of each POS tag was calculated as a percentage of total tokens in the transcript. To quantify the divergence from ground truth, we computed the following difference score for each POS category: Diff = ASR% − GT% where ASR% denotes the normalized frequency of a given POS tag in the ASR output, and GT% refers to the same in the human-annotated ground truth. This method captures low-level grammatical discrepancies that may not be evident through surface-level metrics such as Word Error Rate (WER). By analyzing syntactic token distributions, we can detect whether the ASR system correctly retains functional elements like auxiliaries, determiners, and prepositions, which are often critical in clinical communication. However, while POS distribution analysis highlights structural preservation, it abstracts away semantic meaning thus, semantically divergent but syntactically similar sentences may yield indistinguishable POS profiles. Furthermore, tagging accuracy is inherently dependent on transcript quality suggesting POS distribution analysis is best used in conjunction with other metrics to form a more comprehensive evaluation of ASR performance. Part-of-Speech (POS) Sequential Word Error Rate (POS-sqWER) The Part-of-Speech Sequential Word Error Rate ( POS-sqWER ) is a syntactically motivated evaluation metric that measures ASR fidelity by comparing the sequence of POS tags in the ASR output to the corresponding ground truth sequence. Each transcript was first POS-tagged using a standard NLP parser (Stanza) and then aligned against its reference using Levenshtein distance 22 . The following equation is applied after realignment has occurred. $$\:POSsqWER=\frac{{S}_{\:}+\:D\:+\:I}{{N}_{POS}}\times\:100\:$$ Like WER, the above formula computes the minimum number of insertions (I), deletions (D), and substitutions (S) needed to transform the ASR-derived POS tag sequence into the reference POS tag sequence after alignment and it is normalized by N POS, the number of POS tokens in the reference sequence. By focusing on grammatical token types instead of surface words, POS-sqWER focuses on syntactic disruptions such as omission of functional words (e.g., prepositions, determiners), and reordering of syntactic units (e.g., misplaced conjunctions). This syntactic-level evaluation enables comparisons across ASR systems, especially in populations where syntactic degradation is a clinical marker (e.g., dementia, aphasia). As such, POS-sqWER complements traditional WER by offering a fine-grained lens on grammatical fidelity, supporting more informed model selection for clinical NLP and diagnostic applications. EXAMPLE: Ground Truth (GT): “The patient is experiencing frequent nausea and occasional vomiting.” ASR Output: “The patients experiencing frequent noises and occasional bombing.” [Insert Table 3 here] Table 3 Example sentence and POS analysis. Word (GT) POS (GT) Word (ASR) POS (ASR) The DET The DET patient NOUN patients NOUN is AUX — (dropped) — experiencing VERB experiencing VERB frequent ADJ frequent ADJ nausea NOUN noises NOUN and CCONJ and CCONJ occasional ADJ occasional ADJ vomiting NOUN bombing NOUN WER = (3 + 1 + 0) / 9 = 0.44 POS-sqWER = (0 + 1 + 0) / 9 = 0.11 Table 3 illustrates an example in which WER is high while POS-sqWER remains lower, as all word substitutions occur within the same part-of-speech category as the ground truth. As a result, the grammatical structure of the ASR transcript is preserved, even though the semantic content of the sentence is substantially altered. This distinction highlights how POS-sqWER captures differences between grammatical consistency and semantic distortion that are not reflected by standard WER. Stutter Detection Error (SDE) We quantified stutter detection using paired ground-truth (GT) and ASR transcripts, restricting analysis to sentences that contained ≥ 1 GT stutter (sentences without GT stutters were excluded). Human annotators marked GT stutters with hyphens within or between tokens (e.g., “i-i-is”, “t- thing”, “ca- casey”), and we preserved hyphens during preprocessing. ASR outputs, by contrast, did not use hyphens; when stutters were present, they appeared as repeated tokens or character elongations (e.g., “i i i is”, “ttthing”, “c casey”). We operationalized a GT stutter event as any partial word immediately followed by a hyphen, with or without an intervening space before the completed word. For each GT event we extracted the target word (e.g., “ca- casey” → “casey”; “i-i-is” → “is”) and searched the aligned ASR sentence for a corresponding stutter realization of the same target, accepting token repeats or character elongations. To accommodate minor orthographic variation, matches were confirmed with fuzzy lexical matching (threshold = 90). We maintained an exclusion list for surface partials that occur fluently without reflecting a stutter (e.g., “we went”, “up upon”, “for forty”, “an animal”, “we were”, “not noticing”, “was washing”), which were not credited as detections. When ASR produced more stutter-like forms than GT, only matches to GT targets were counted as correct (extras were ignored). For each sentence, we computed trials = number of GT stutter events, correct = number matched in ASR, errors = trials − correct, and defined SDE = errors / trials (Accuracy = 1 − SDE). Sentence Embedding Distance Analysis Using Cosine Distance To assess the semantic fidelity of ASR outputs beyond lexical or syntactic similarity, we employed a sentence embedding-based evaluation approach. This method maps both the ASR transcription and the corresponding ground truth sentences into high-dimensional semantic representations, allowing for a quantitative comparison of their meanings. In this study, we used the "all-MiniLM-L6-v2" model from Huggingface’s SentenceTransformers library 45 , 46 , a transformer-based sentence encoder. This model is trained via knowledge distillation to approximate the performance of larger models while maintaining high efficiency, making it particularly suitable for real-time or large-scale clinical applications 46 . Each transcript, both ground truth and ASR-generated was encoded into a 384-dimensional embedding. The cosine distance between the embeddings of each ASR sentence and its ground truth sentence was computed providing a semantic similarity score. This approach captures similarities between paraphrased or restructured content that traditional word-level metrics like WER or POS-based scores will always categorize as errors. For example, while WER penalizes synonym substitutions or reordered clauses, sentence embeddings can tolerate such variations if the overall meaning remains intact. Statistical methods We analysed five metrics (WER, POS-sqWER, POS_MDE, SDE, and cosine distance) using mixed-effects models to account for repeated measures by recording. For WER and POS-sqWER, we fit linear mixed-effects models with fixed effects for ASR Model, Group (HC vs DEM), and their interaction, and a random intercept for Filename . To stabilise variance and accommodate zeros, outcomes were modelled on the log1p scale and then back-transformed to percentages (i.e., \(\:100\times\:(\text{e}\text{x}\text{p}(\widehat{\mu\:})-1)\) ). Extreme values were trimmed at 150% prior to modelling to reduce undue influence of outliers. We report model-estimated marginal means (“adjusted” values) and 95% CIs per ASR model × Group, with Tukey-adjusted pairwise comparisons. For POS_MDE, we fit a linear mixed-effects model with fixed effects for ASR Model, Group, POS Tag (13 levels), and all two- and three-way interactions, plus a random intercept for Filename . POS_MDE was analysed on the log1p scale and back-transformed for presentation. We examined planned, Tukey-adjusted pairwise contrasts within Group × POS Tag to localise differences among ASR models. For SDE, we used a binomial GLMM on sentence-level counts (errors vs. Correct), with fixed effects for ASR Model, Group, and their interaction, and a random intercept for Filename when supported by the data. Results are reported as estimated marginal means on the response scale (with 95% CIs) and as Tukey/Holm-adjusted odds-ratio contrasts among ASR models. For cosine distance, we fit a linear mixed-effects model with ASR Model, Group, and their interaction as fixed effects and a random intercept for Filename . Lower values indicate better semantic fidelity; adjusted means and Tukey-adjusted pairwise tests were computed on the raw scale. Across models, we verified standard assumptions: approximate normality and homoscedasticity of residuals on the transformed scale for lmms; binomial variance and absence of complete separation for the GLMM; and conditional independence given random effects. Denominator degrees of freedom used Satterthwaite or Kenward–Roger approximations where applicable. Estimated marginal means and multiplicity-controlled comparisons were obtained from emmeans 47 . Analyses were conducted in R (e.g., lme4/lmertest 48 , glmmtmb 49 ). Results To evaluate the transcription accuracy of each ASR system, we computed the WER by aligning each ASR-generated transcript with its corresponding ground truth transcript. [Insert Fig. 2 here] Figure 2 shows that the model-adjusted WER had a strong main effect of ASR model \(\:\left[F\right(\text{7,2106.47})=290.03,p<0.001]\:\) and Group \(\:\left[F\right(\text{1,303.09})=152.53,p<0.001]\) , and a significant ASR Model × Group interaction \(\:\left[F\right(\text{7,2106.47})=16.34,p<0.001]\) . Given the interaction, we examined ASR differences within each group using Tukey-adjusted pairwise comparisons. Within DEM, Google had the highest adjusted WER and was significantly worse than every other system (all \(\:p<0.001\) ); Meta and Microsoft were also significantly lower than the top performers (all \(\:p\le\:0.001\) ). A leading cluster consisted of CrisperWhisper, AWS, AssemblyAI, and Nemo: CrisperWhisper significantly (WER=16.1) outperformed AssemblyAI and Nemo (both \(\:p\le\:0.001\) ) and AWS was statistically indistinguishable from AssemblyAI and Nemo ( \(\:p=0.24\) and \(\:p=0.40\) , respectively). Within HC, Google again had the highest adjusted WER and was significantly lower than all others (all \(\:p<0.001\) ); Meta was also significantly worse than the better systems (all \(\:p<0.001\) ). The lowest WERs were obtained by AWS (WER=6.94) and CrisperWhisper (WER=6.76), both outperforming Microsoft, Nemo, and Whisper (all \(\:p\le\:0.006\) ). Taken together, these results indicate that performance rankings vary by group, with Google and Meta exhibiting significantly higher error rates, and a top tier formed by CrisperWhisper and AWS across both groups, alongside AssemblyAI and Nemo in DEM and AssemblyAI, Microsoft, Nemo and Whisper forming a statistically similar mid-tier in HC. [Insert Fig. 3 here] Figure 3 shows the Model-adjusted POS-sqWERs showed strong main effects of ASR Model \(\:\left[F\right(\text{7,2128})=150.15,p<0.001]\:\) and Group \(\:\left[F\right(\text{1,304})=153.72,p<0.001]\) , and a significant ASR Model × Group interaction \(\:\left[F\right(\text{7,2128})=19.46,p<0.001]\) . Given the interaction, we compared ASRs within each group using Tukey adjustment. Within DEM, Google and Microsoft had the highest POS-sqWER and were indistinguishable \(\:(p=1.00)\) , each significantly worse than AWS, CrisperWhisper, Nemo, AssemblyAI, and Whisper (all \(\:\text{p}\text{}\text{<}\text{}0.001)\) . Meta was also worse than the better-performing systems (all \(\:\text{p}\text{}\text{<}\text{}0.001)\) . A leading cluster comprised AWS and CrisperWhisper (no difference, \(\:p=0.997\) ); Nemo did not differ significantly from AWS \(\:\left(p=0.30\right)\:\) or CrisperWhisper \(\:(p=0.056)\) . Within HC, Google again had the highest POS-sqWER, significantly worse than every other ASR (all \(\:\text{p<}0.001)\) . Meta was also high and worse than AWS, CrisperWhisper, Microsoft, Nemo, and Whisper (all \(\:\text{p}\text{}\text{≤}\text{}0.001)\) . AWS and CrisperWhisper formed the top tier with no difference between them \(\:(p=0.999)\) . Overall, across groups, AWS and CrisperWhisper consistently yielded the lowest adjusted POS-sqWER, as with WER. Nemo performed next best, followed by AssemblyAI and Whisper. Microsoft had middling performance with health control samples but had almost the worst performance for samples from participants with dementia. Google and Meta exhibited the highest POS-sqWER, with Google repeatedly the worst. [Insert Fig. 4 here] Figure 4 provides a graphical overview of the POS_MDE results. There were robust main effects of ASR model, Group, and POS tag: ASR_Model, F(7, 27,243.0) = 68.06, p < 2.2×10⁻¹⁶; Group, F(1, 297.4) = 27.25, p = 3.36×10⁻⁷; POS_Tag, F(12, 27,251.1) = 82.19, p < 2.2×10⁻¹⁶. Two-way interactions were also significant: ASR_Model × Group, F(7, 27,243.0) = 10.51, p = 2.97×10⁻¹³; ASR_Model × POS_Tag, F(84, 27,242.9) = 1.60, p = 4.44×10⁻⁴; Group × POS_Tag, F(12, 27,251.1) = 23.29, p < 2.2×10⁻¹⁶. The three-way interaction ASR_Model × Group × POS_Tag was not significant, F(84, 27,242.9) = 0.91, p = 0.707. Taken together, these effects indicate (i) overall differences in POS_MDE among ASR systems, (ii) higher POS_MDE in DEM vs HC on average, and (iii) substantial tag-specific variation, with the pattern of ASR differences varying by group and POS tag. Planned, Tukey-adjusted pairwise comparisons within Group × POS_Tag show where models differ. Illustrative contrasts (negative estimates indicate the first model has lower POS_MDE than the second): Adjectives (ADJ): differences were small; only a few contrasts reached significance (e.g., DEM: AssemblyAI < Google, estimate = − 0.0086, p = 0.0045; AWS < Google, − 0.0077, p = 0.0169). Adpositions (ADP): several models exhibited lower POS_MDE than Google (e.g., DEM: Microsoft < Google, − 0.0096, p = 0.0005; HC: multiple models vs Google trended lower, several p ≤ 0.0065). Pronouns (PRON) and Nouns (NOUN): consistent and sizable gaps favored multiple systems over Google in both groups (e.g., DEM PRON: AssemblyAI < Google, − 0.0176, p < 0.0001; AWS < Google, − 0.0114, p < 0.0001; DEM NOUN: AssemblyAI < Google, − 0.0224, p < 0.0001; CrisperWhisper < Google, − 0.0188, p < 0.0001). Verbs (VERB): again, many models showed lower POS_MDE than Google (e.g., DEM: AssemblyAI < Google, − 0.0107, p < 0.0001; CrisperWhisper < Google, − 0.0122, p < 0.0001). Interjections (INTJ): Meta was notably higher POS_MDE than several models in DEM (e.g., DEM: AssemblyAI < Meta, − 0.0144, p < 0.0001; AWS < Meta, − 0.0189, p < 0.0001), while most differences in HC were not significant. Determiners (DET): Microsoft showed higher POS_MDE than many peers in DEM (e.g., DEM: AWS < Microsoft, − 0.0156, p < 0.0001; CrisperWhisper < Microsoft, − 0.0166, p < 0.0001), whereas differences in HC were largely nonsignificant. Because lower POS_MDE indicates a closer match between ASR-predicted and human POS distributions, the above contrasts suggest (a) DEM speech increases POS distributional mismatch relative to HC; (b) some systems (e.g., Google in several tags; Microsoft for DET in DEM) exhibit larger mismatches; and (c) the ranking is tag-dependent, underscoring the value of POS-resolved evaluation. [Insert Fig. 5 here] SDE was modelled with a binomial GLMM including ASR Model, Group (HC vs DEM), and their interaction. There was a significant main effect of ASR Model (Type-III Wald χ²(7) = 23.76, p = 0.0013), no main effect of Group (χ²(1) = 0.31, p = 0.581), and no ASR×Group interaction (χ²(7) = 4.47, p = 0.724). Pairwise differences are reported as odds ratios (OR). For a contrast written as ‘Model1 / Model2,’ OR 1 means Model1 has higher odds of SDE than Model2. Within DEM, several contrasts were significant after Tukey adjustment (Fig. 5 ). Notably, CrisperWhisper (SDE = 62.76) showed substantially lower SDE than multiple models, for example, CrisperWhisper vs Google OR = 0.19, p <.0001; CrisperWhisper vs Meta OR = 0.19, p <.0001; CrisperWhisper vs Microsoft OR = 0.28, p <.0001; and CrisperWhisper vs Whisper OR = 0.11, p <.0001. AWS also had lower SDE than Whisper (AWS vs Whisper OR = 0.24, p <.0001). By contrast, AssemblyAI exhibited higher SDE than several peers (e.g., AssemblyAI vs AWS OR = 4.04, p <.0001; AssemblyAI vs CrisperWhisper OR = 8.49, p <.0001; AssemblyAI vs Nemo OR = 3.93, p = 0.0001). Differences among Google, Meta, and Microsoft were generally not significant (e.g., Google vs Meta OR ≈ 1.00, p = 1.00). Within HC, no pairwise comparison reached significance after Tukey adjustment (all p ≥ 0.466), consistent with the non-significant main effect of Group and the null ASR×Group interaction. Taken together, SDE varied by ASR Model overall, with CrisperWhisper (and often AWS) showing lower odds of SDE than several alternatives in the DEM group, while Whisper and AssemblyAI tended to fare worse in those same comparisons. No reliable ASR ordering emerged within HC. [Insert Fig. 6 here] A linear mixed-effects model (ASR_Model × Group; random intercept for Filename) showed strong main effects of ASR model and Group, and a significant interaction (ASR_Model: \(\:F\left(\text{7,2128}\right)=300.31,p<2.2\times\:{10}^{-16}\) ; Group: \(\:F\left(\text{1,323.4}\right)=60.31,p=1.08\times\:{10}^{-13}\) ; ASR_Model×Group: \(\:F\left(\text{7,2128}\right)=11.23,p=4.93\times\:{10}^{-14}\) ). Within-group Tukey comparisons (on the raw cosine-distance scale; lower is better) indicated that Google had substantially higher (worse) cosine distances than every other ASR in both groups (all \(\:p<0.0001;\text{F}\text{i}\text{g}.\:6\) ). In DEM, AssemblyAI, CrisperWhisper, Whisper, and (to a lesser extent) AWS formed the leading tier: AssemblyAI outperformed AWS and Microsoft (both \(\:p\le\:0.0012\) ) and was statistically indistinguishable from CrisperWhisper and Whisper ( \(\:p\ge\:0.93\) ); CrisperWhisper and Whisper were also indistinguishable ( \(\:p=1.00\) ) and each beat AWS ( \(\:p\le\:0.035\) ). Meta and Microsoft were reliably worse than this leading tier but still markedly better than Google (all contrasts vs Google \(\:p<0.0001\) ). Nemo sat mid-pack—better than Google ( \(\:p<0.0001\) ) but worse than AssemblyAI ( \(\:p=0.033\) ). In HC, the same pattern held: Google was again worse than all peers (all \(\:p<0.0001\) ); CrisperWhisper and Whisper were statistically similar ( \(\:p=1.00\) ), AWS was similar to Whisper ( \(\:p=1.00\) ), and Microsoft and Nemo were indistinguishable ( \(\:p=1.00\) ). Collectively, these results confirm meaningful model-wise differences in semantic fidelity (cosine distance) that vary by Group, with Google consistently underperforming and a top cluster comprising Whisper, CrisperWhisper, AssemblyAI (and often AWS) exhibiting the lowest distances. Discussion Our multi-metric evaluation highlights that model choice and model robustness depend strongly on both the linguistic target and speaker group. Across global accuracy measures, WER and POS-sqWER produced similar trends: performance was markedly better for HC than DEM, and ASR models differed substantially. In both groups Google exhibited consistently higher error rates, while AWS and CrisperWhisper formed a reliable top tier, joined by AssemblyAI and Nemo in DEM and by several mid-tier contenders (AssemblyAI, Microsoft, Nemo and Whisper) in HC. This HC–DEM gap, replicated again in the cosine-distance analysis of sentence embeddings, underscores a persistent vulnerability of current systems to atypical prosody and articulation an effect likely rooted in training corpora skewed toward normative speech. Beyond token-level accuracy, the POS_MDE analysis shows why “which model is best” is intrinsically task-dependent. There were strong main effects of ASR, Group, and POS tag, and the pattern of differences shifted across parts of speech: several models showed large mismatches for specific categories in DEM (e.g., Google across multiple tags; Microsoft for determiners), whereas other categories showed only small, scattered differences. Nemo was frequently among the strongest at preserving POS distributions, and Whisper and AssemblyAI were broadly competitive except for isolated tags (e.g., VERB for Whisper, SCONJ for AssemblyAI in DEM). These tag-resolved effects argue for selecting models with an eye to the linguistic constructs most critical for a given downstream analysis rather than relying on a single omnibus score. Stutter detection exposed a different facet of system behaviour. Although there was a significant overall ASR effect, there was no group effect and no ASR×Group interaction, and absolute accuracies were low for all models. Within DEM, CrisperWhisper exhibited reliably lower odds of stutter-detection error than several competitors; in HC, pairwise differences were not reliable, consistent with the very small number of stutter events in that group. The uniformly low performance here indicates that disfluency modelling remains an unmet need: even systems that are accurate on words and semantics struggle to capture clinically salient disfluencies that often drive assessment decisions. Semantic fidelity (cosine distance) reinforced the above conclusions while adding nuance. Again, HC transcripts were closer to the ground truth than DEM, and there were pronounced model effects with a significant interaction by Group. Google had the largest (worst) cosine distances in both groups by a wide margin. A leading cluster Whisper, CrisperWhisper, and AssemblyAI, often joined by AWS showed the smallest distances, with Meta and Microsoft clearly behind that cluster but still substantially better than Google. Nemo tended to sit mid-pack on this measure despite its strong POS behaviour. Together with WER and POS-sqWER, these findings suggest that models can trade off token-level accuracy and sentence-level semantic alignment in subtle ways; a model that is mid-tier on WER can still preserve meanings well, and vice-versa. Practically, three themes emerge. First, group matters: every metric showed a degradation for DEM, signaling that off-the-shelf models need adaptation or fine-tuning to handle atypical speech. Second, metric choice matters: AWS and CrisperWhisper are consistently strong on WER and POS-sqWER; Whisper, CrisperWhisper, and AssemblyAI often lead on semantic alignment. Third, use-case matters: if the endpoint depends on specific grammatical categories (e.g., function-word distributions), Nemo or Whisper and AssemblyAI may be preferable; for general clinical transcription where token accuracy dominates, AWS and CrisperWhisper are safe defaults; for workflows requiring detection of disfluency, none of the models yet meet a clinically useful bar, with CrisperWhisper the current but limited front-runner in DEM. Our study’s strength is its unified framework spanning token errors, syntactic structure, semantics, and disfluency, revealing that “accuracy” is multi-dimensional and that rankings shift when the lens changes. Limitations include relatively sparse stutter events in HC and the possibility that some POS-resolved differences reflect sampling variability at the tag level; larger, more diverse clinical corpora and model adaptation to pathological speech should narrow uncertainty bands and may alter rankings. Still, the overall pattern is clear: modern ASR systems can perform well on normative speech yet manifest predictable failure modes in clinical contexts. Model selection should therefore be purpose-built, and future development should prioritize robustness to atypical speech and explicit modelling of disfluencies if ASR is to support reliable, fair, and clinically useful inference. Limitations and future research Our dataset includes linguistic samples from a group of Australian healthy older adults and adults with dementia. Thus, we cannot generalize our findings to broader populations, such as children or younger adults, speakers of other languages, or other clinical conditions such as Parkinson’s disease, ALS, or developmental speech impairments. Notably, the analysis of stutter detection was constrained by the markedly low incidence of stutters in the healthy group (only 11 instances), which warrants exploration of these ASR models using language data from a group of speakers who experience stuttering or other disfluencies to evaluate their performance. Although our evaluation spans eight ASR systems covering commercial, open-source, and fine-tuned models, this set is not exhaustive, and the inclusion of proprietary or next-generation models might yield different outcomes. Second, the syntactic and semantic analyses employed rely on general-purpose NLP models (in our case, Stanza) for POS tagging and all-MiniLM-L6-v2 for sentence embeddings. These models are not specifically trained on clinical speech and may inadequately capture the nuances of pathological language, such as irregular grammar, disfluencies, or atypical prosody. As a result, transcription errors with clinical significance, particularly those involving disfluency or language impairment, may be underestimated or mis-characterised. Third, while POS-sqWER introduces valuable grammatical sensitivity into ASR evaluation, it currently assigns equal weight to all part-of-speech mismatches. This uniform weighting overlooks the differing functional and diagnostic significance of various POS categories. For example, errors in nouns or verbs may carry more clinical consequence than those involving function words like determiners or auxiliary verbs. Similarly, sentence embedding-based semantic metrics, though effective for capturing general meaning preservation, may miss fine-grained deviations that are critical in clinical interpretation, such as missed negations, disfluent insertions, or altered temporal references. Our results demonstrate areas of future research in the improvement in both ASR systems and their evaluation. Expanding training datasets for ASR models to include a more balanced representation of speech, language and neurological conditions could improve available ASR models. In terms of testing and refining these models, incorporating a wide suite of performance metrics (as we have done here) could significantly enhance the diagnostic relevance and precision of ASR technologies in uses involving clinical populations. Conclusion Across metrics and groups, AWS and CrisperWhisper were the most consistently strong performers, while Google and Meta underperformed. AssemblyAI, Nemo, Microsoft, and Whisper generally occupied a mid-tier with performance varying by metric and, in some cases, by linguistic category. Participants with dementia were transcribed less accurately than healthy controls across nearly all measures, reinforcing the clinical gap that remains for atypical speech. POS-MDE highlighted tag-specific differences that can inform model choice when particular parts of speech matter for a given application. Stutter detection was weak for all systems, although CrisperWhisper was comparatively better in the dementia group and absolute accuracies were still low. Taken together, a multi-metric evaluation provides a more faithful and actionable picture of clinical ASR performance than reliance on WER alone. Declarations Declarations of interest Drs Kathiresan and Profs Vogel, and Ms. Gasparini are employees of Redenlab Ltd., a speech neuroscience company. Ms. Zona Ling was an intern at Redenlab during this work. Acknowledgements We acknowledge funding support through the CUREator+ Dementia and Cognitive Decline BioMedTech Incubator program supported by ANDHealth and partners, which contributed to the digital health research environment underpinning this work. Author contributions: TK: Conceptualization, Data curation, Formal analysis, Methodology, Project administration, Visualization, Writing - original draft, Writing - review & editing ZL: Data curation, Methodology, Software, Writing - review & editing LG: Data curation, Writing - review & editing SEP: Data curation, Writing - review & editing CT: Data curation, Writing - review & editing HE: Data curation, Writing - review & editing AB: Methodology, Writing - review & editing APV: Conceptualization, Funding acquisition, Methodology, Supervision, Writing - review & editing Data sharing statement: All data, code, and results are available at:https://github.com/kathiresant-redenlab/ASR_MultiMetric_Evaluation Declaration of generative AI and AI-assisted technologies in the manuscript preparation process During the preparation of this work, authors used Large Language Models (ChatGPT) to edit their writing for grammatical correctness and clarity. After using these tools, the authors reviewed and edited the content and take full responsibility for the content of the published article. References Cornell, S. et al. The CHiME-7 DASR Challenge: Distant Meeting Transcription with Multiple Devices in Diverse Scenarios. Preprint at https://doi.org/10.48550/ARXIV.2306.13734 (2023). Shah, M. A., Noguero, D. S., Heikkila, M. A., Raj, B. & Kourtellis, N. Speech Robust Bench: A Robustness Benchmark For Speech Recognition. Preprint at https://doi.org/10.48550/arXiv.2403.07937 (2024). Jafari, Z., Andrew, M. K. & Rockwood, K. J. Diagnostic utility of speech-based biomarkers in mild cognitive impairment: a systematic review and meta-analysis. Age Ageing 54, afaf316 (2025). Falcetta, F. S., De Almeida, F. K., Lemos, J. C. S., Goldim, J. R. & Da Costa, C. A. Automatic documentation of professional health interactions: A systematic review. Artif. Intell. Med. 137, 102487 (2023). Lin, H. et al. Identification of digital voice biomarkers for cognitive health. Explor. Med. 1, 406–417 (2020). Hecker, P., Steckhan, N., Eyben, F., Schuller, B. W. & Arnrich, B. Voice Analysis for Neurological Disorder Recognition–A Systematic Review and Perspective on Emerging Trends. Front. Digit. Health 4, 842301 (2022). García, A. M., De Leon, J., Tee, B. L., Blasi, D. E. & Gorno-Tempini, M. L. Speech and language markers of neurodegeneration: a call for global equity. Brain 146, 4870–4879 (2023). Cao, F., Vogel, A. P., Gharahkhani, P. & Renteria, M. E. Speech and language biomarkers for Parkinson’s disease prediction, early diagnosis and progression. Npj Park. Dis. 11, 57 (2025). Chan, J. C. S., Stout, J. C. & Vogel, A. P. Speech in prodromal and symptomatic Huntington’s disease as a model of measuring onset and progression in dominantly inherited neurodegenerative diseases. Neurosci. Biobehav. Rev. 107, 450–460 (2019). Noffs, G. et al. What speech can tell us: A systematic review of dysarthria characteristics in Multiple Sclerosis. Autoimmun. Rev. 17, 1202–1209 (2018). Green, J. R. et al. Bulbar and speech motor assessment in ALS: Challenges and future directions. Amyotroph. Lateral Scler. Front. Degener. 14, 494–500 (2013). Poole, M. L., Brodtmann, A., Darby, D. & Vogel, A. P. Motor Speech Phenotypes of Frontotemporal Dementia, Primary Progressive Aphasia, and Progressive Apraxia of Speech. J. Speech Lang. Hear. Res. 60, 897–911 (2017). Duffy, J. R., Utianski, R. L. & Josephs, K. A. Primary progressive apraxia of speech: from recognition to diagnosis and care. Aphasiology 35, 560–591 (2021). Croot, K., Ballard, K., Leyton, C. E. & Hodges, J. R. Apraxia of Speech and Phonological Errors in the Diagnosis of Nonfluent/Agrammatic and Logopenic Variants of Primary Progressive Aphasia. J. Speech Lang. Hear. Res. 55, (2012). Leung, W.-Z., Cross, M., Ragni, A. & Goetze, S. Training Data Augmentation for Dysarthric Automatic Speech Recognition by Text-to-Dysarthric-Speech Synthesis. in Interspeech 2024 2494–2498 (ISCA, 2024). doi: 10.21437/Interspeech.2024-1645 . Zhou, L. et al. Analysis of Errors in Dictated Clinical Documents Assisted by Speech Recognition Software and Professional Transcriptionists. JAMA Netw. Open 1, e180530 (2018). Van Buchem, M. M. et al. The digital scribe in clinical practice: a scoping review and research agenda. Npj Digit. Med. 4, 57 (2021). Teleki, M., Dong, X., Kim, S. & Caverlee, J. Comparing ASR Systems in the Context of Speech Disfluencies. in Interspeech 2024 4548–4552 (ISCA, 2024). doi: 10.21437/Interspeech.2024-1270 . Michot, J. et al. Error-preserving Automatic Speech Recognition of Young English Learners’ Language. in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 6444–6454 (Association for Computational Linguistics, Bangkok, Thailand, 2024). doi: 10.18653/v1/2024.acl-long.348 . Radford, A. et al. Robust Speech Recognition via Large-Scale Weak Supervision. Preprint at https://doi.org/10.48550/ARXIV.2212.04356 (2022). Pratap, V. et al. Scaling Speech Technology to 1,000 + Languages. Preprint at https://doi.org/10.48550/ARXIV.2305.13516 (2023). Navarro, G. A guided tour to approximate string matching. ACM Comput. Surv. 33, 31–88 (2001). Tran, B. D., Mangu, R., Tai-Seale, M., Lafata, J. E. & Zheng, K. Automatic speech recognition performance for digital scribes: a performance comparison between general-purpose and specialized models tuned for patient-clinician conversations. AMIA Annu. Symp. Proc. AMIA Symp. 2022, 1072–1080 (2022). Mirheidari, B., Blackburn, D., Walker, T., Reuber, M. & Christensen, H. Dementia detection using automatic analysis of conversations. Comput. Speech Lang. 53, 65–79 (2019). Miner, A. S. et al. Assessing the accuracy of automatic speech recognition for psychotherapy. Npj Digit. Med. 3, 82 (2020). Schultz, B. G. et al. Automatic speech recognition in neurodegenerative disease. Int. J. Speech Technol. 24, 771–779 (2021). Afonja, T. et al. Performant ASR Models for Medical Entities in Accented Speech. in Interspeech 2024 2315–2319 (ISCA, 2024). doi: 10.21437/Interspeech.2024-2261 . Le-Duc, K. et al. Medical Spoken Named Entity Recognition. in Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track) 724–783 (Association for Computational Linguistics, Albuquerque, New Mexico, 2025). doi: 10.18653/v1/2025.naacl-industry.59 . Shapira, O., Chazan, S. & Cohen, A. D. N. Measuring the Effect of Transcription Noise on Downstream Language Understanding Tasks. in Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 29978–30004 (Association for Computational Linguistics, Vienna, Austria, 2025). doi: 10.18653/v1/2025.acl-long.1449 . Cao, J. et al. A Comparative Analysis of Automatic Speech Recognition Errors in Small Group Classroom Discourse. in Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization 250–262 (ACM, Limassol Cyprus, 2023). doi: 10.1145/3565472.3595606 . Amazon Web Services, Inc. Amazon Transcribe. (2025). https://aws.amazon.com/transcribe Hugging Face. nvidia/stt_en_conformer_ctc_large. (2025). https://huggingface.co/nvidia/stt_en_conformer_ctc_large AssemblyAI. AssemblyAI API Documentation . (2024). https://www.assemblyai.com/docs Microsoft. microsoft/Phi-4-multimodal-instruct. (2025). https://huggingface.co/microsoft/Phi-4-multimodal-instruct Wagner, L., Thallinger, B. & Zusag, M. CrisperWhisper: Accurate Timestamps on Verbatim Speech Transcriptions. Preprint at https://doi.org/10.48550/arXiv.2408.16589 (2024). Nyra Health. CrisperWhisper: A custom automatic speech recognition model based on Whisper. GitHub repository (2024). https://github.com/nyrahealth/CrisperWhisper Pratap, V. et al. Scaling Speech Technology to 1,000 + Languages. Preprint at https://doi.org/10.48550/arXiv.2305.13516 (2023). Google Cloud. Recognize Speech by Using Medical Models (Speech-to-Text V1) . (2025). Language Development Project Team. 4. Utterance boundaries. (2015). Kahane, S., Caron, B., Strickland, E. & Gerdes, K. Annotation guidelines of UD and SUD treebanks for spoken corpora: A proposal. in Proceedings of the 20th International Workshop on Treebanks and Linguistic Theories (TLT, SyntaxFest 2021) (eds Dakota, D., Evang, K. & Kübler, S.) 35–47 (Association for Computational Linguistics, Sofia, Bulgaria, 2021). MacWhinney, B. CHAT Manual. Preprint at https://doi.org/10.21415/3MHN-0Z89 (2019). MacWhinney, B. & Fromm, D. Language Sample Analysis With TalkBank: An Update and Review. Front. Commun. 7, 865498 (2022). National Institute of Standards and Technology. Speech Recognition Scoring Toolkit (SCTK) and \textttsclite. (2021). Qi, P., Zhang, Y., Zhang, Y., Bolton, J. & Manning, C. D. Stanza: A Python Natural Language Processing Toolkit for Many Human Languages. in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations 101–108 (Association for Computational Linguistics, Online, 2020). doi: 10.18653/v1/2020.acl-demos.14 . Reimers, N. & Gurevych, I. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) 3982–3992 (2019). doi: 10.18653/v1/D19-1410 . Reimers, N. & Gurevych, I. Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation. in Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) 4512–4525 (Association for Computational Linguistics, Online, 2020). doi: 10.18653/v1/2020.emnlp-main.365 . Searle, S. R., Speed, F. M. & Milliken, G. A. Population Marginal Means in the Linear Model: An Alternative to Least Squares Means. Am. Stat. 34, 216–221 (1980). Lenth, R. V. Least-Squares Means: The R Package lsmeans. J. Stat. Softw. 69, (2016). Brooks, M., E. et al. glmmTMB Balances Speed and Flexibility Among Packages for Zero-inflated Generalized Linear Mixed Modeling. R J. 9, 3 Additional Declarations Competing interest reported. Drs Kathiresan and Vogel, and Ms Gasparini, are employees of Redenlab Ltd., a speech neuroscience company. Ms Zona Ling was an intern at Redenlab Ltd. during the course of this work. Supplementary Files SupplementaryMaterials.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8824529","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":591462295,"identity":"eda1151b-cbc2-4967-86f1-4574eade137d","order_by":0,"name":"Thayabaran Kathiresan","email":"data:image/png;base64,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","orcid":"","institution":"University of Melbourne","correspondingAuthor":true,"prefix":"","firstName":"Thayabaran","middleName":"","lastName":"Kathiresan","suffix":""},{"id":591462296,"identity":"b95d4266-1b5b-4b1a-9b5e-26e820e54115","order_by":1,"name":"Zona Ling","email":"","orcid":"","institution":"Redenlab Ltd","correspondingAuthor":false,"prefix":"","firstName":"Zona","middleName":"","lastName":"Ling","suffix":""},{"id":591462298,"identity":"58ce5fe6-3901-40c3-9f2c-0fa21056984e","order_by":2,"name":"Loretta Gasparini","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Loretta","middleName":"","lastName":"Gasparini","suffix":""},{"id":591462302,"identity":"0719ab81-4915-4ec6-aa8c-746f4334426c","order_by":3,"name":"Stephanie Eliza Piccolo","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Stephanie","middleName":"Eliza","lastName":"Piccolo","suffix":""},{"id":591462307,"identity":"93d52c6f-22ec-41b0-ad25-1eafdfb3dc0b","order_by":4,"name":"Clara Tjiandra","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Clara","middleName":"","lastName":"Tjiandra","suffix":""},{"id":591462311,"identity":"13b8c9f8-3c1e-4b04-b99c-b1e480a830c5","order_by":5,"name":"Hana Elhammamy","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Hana","middleName":"","lastName":"Elhammamy","suffix":""},{"id":591462312,"identity":"dde5485b-07e2-43f9-9c86-ddde3516d3cc","order_by":6,"name":"Amy Brodtmann","email":"","orcid":"","institution":"Monash University","correspondingAuthor":false,"prefix":"","firstName":"Amy","middleName":"","lastName":"Brodtmann","suffix":""},{"id":591462313,"identity":"5475377a-4487-4bcd-a05c-8a9106890665","order_by":7,"name":"Adam P. Vogel","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Adam","middleName":"P.","lastName":"Vogel","suffix":""}],"badges":[],"createdAt":"2026-02-08 23:53:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8824529/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8824529/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102818346,"identity":"e5ad7a9a-c914-4315-bd88-9ebc129ac484","added_by":"auto","created_at":"2026-02-17 06:43:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":277882,"visible":true,"origin":"","legend":"\u003cp\u003eEvaluation framework for benchmarking ASR performance in dementia and healthy older adult speech.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8824529/v1/20f4187d73067029170fb343.png"},{"id":102818344,"identity":"b4a7d903-c00d-439a-980d-d34a60998b0f","added_by":"auto","created_at":"2026-02-17 06:43:40","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":37698,"visible":true,"origin":"","legend":"\u003cp\u003eComparative adjusted word error rate (WER) of ASR models in healthy controls (HC) and individuals with dementia (DEM).\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8824529/v1/f24435449b1bbbc04b37acc0.jpeg"},{"id":102818350,"identity":"2969270b-4c0c-4d10-a10a-0f434a935b24","added_by":"auto","created_at":"2026-02-17 06:43:40","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":37488,"visible":true,"origin":"","legend":"\u003cp\u003eComparative adjusted POS-sqWER of ASR models in HC and DEM.\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8824529/v1/3011cdee4dead1fb70ba4686.jpeg"},{"id":102963576,"identity":"dd53ec9f-e88c-4e96-b9a2-44464d8bf600","added_by":"auto","created_at":"2026-02-19 04:19:05","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":188766,"visible":true,"origin":"","legend":"\u003cp\u003eAdjusted POS_MDE by ASR and Group (HC, DEM) across POS tags.\u003c/p\u003e","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8824529/v1/c821104741ca017183e43098.jpeg"},{"id":102962918,"identity":"dafe20fb-b0c7-4707-806c-289490d46d57","added_by":"auto","created_at":"2026-02-19 04:12:09","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":43833,"visible":true,"origin":"","legend":"\u003cp\u003eModel-adjusted Stutter Detection Error (SDE) across ASR models, stratified by Group (HC, DEM)\u003c/p\u003e","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8824529/v1/3acd293a52e9201c4ec7fff3.jpeg"},{"id":102818348,"identity":"fb501ed1-7857-4f66-b1aa-5a5355f639bc","added_by":"auto","created_at":"2026-02-17 06:43:40","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":41765,"visible":true,"origin":"","legend":"\u003cp\u003eModel-Adjusted Cosine distance across ASR models, stratified by Group (HC, DEM)\u003c/p\u003e","description":"","filename":"floatimage8.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8824529/v1/a79d66aab6e937d243e3f863.jpeg"},{"id":104400549,"identity":"d46794fe-d756-44e9-949b-1d09b8b4e464","added_by":"auto","created_at":"2026-03-11 12:10:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1351468,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8824529/v1/90f48f3c-13e2-47c2-830d-4acfabe99d4d.pdf"},{"id":102818349,"identity":"9c645e9b-aa92-420c-a54c-55552b478086","added_by":"auto","created_at":"2026-02-17 06:43:40","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":16257,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-8824529/v1/4a4620d5e2718fa19bc50b23.docx"}],"financialInterests":"Competing interest reported. Drs Kathiresan and Vogel, and Ms Gasparini, are employees of Redenlab Ltd., a speech neuroscience company. Ms Zona Ling was an intern at Redenlab Ltd. during the course of this work.","formattedTitle":"Beyond word error rate: Multidimensional evaluation of ASR performance for digital speech biomarkers","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAutomatic Speech Recognition (ASR) technologies continue to advance, achieving near-human transcription accuracy in well-controlled settings\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e, showing robustness in adverse, noisy conditions\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. These advances have spurred widespread integration of ASR in consumer applications, enterprise solutions, and in healthcare\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. As digital health ecosystems expand, there is growing interest in leveraging ASR for clinical applications such as remote monitoring, diagnostic support, automated documentation, and telehealth services\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Speech is a rich marker of neurological and cognitive health, offering a non-invasive and accessible clinical assay\u003csup\u003e\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e–\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. However, we do not know how accurately ASR systems process real-world clinical speech, which remains a critical barrier.\u003c/p\u003e \u003cp\u003eUnlike the clean and fluent speech used to train most commercial ASR models, clinical speech is often marked by disfluencies, atypical prosody, articulatory breakdowns, and phonatory disturbances. These features are prominent in individuals with neurodegenerative conditions such as Parkinson’s disease \u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e Huntington’s disease \u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e, multiple sclerosis \u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e, Amyotrophic Lateral Sclerosis (ALS)\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, and various forms of dementia\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e, where speech and language performance varies with disease status. In cases of atypical communication resulting from primary progressive aphasia (PPA) and fronto-temporal dementia, conventional ASR systems may not accurately transcribe words and phrases. PPA and FTD can lead to reduced intelligibility, uncommon phonological patterns, dysfluent or agrammatical phrasing or articulatory imprecision\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, but also to handle ungrammatical utterances. Most available models are trained on clean, non-disordered speech and language, so they tend to produce the most statistically plausible sentence rather than faithfully render atypical productions\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. In this case, rather than transcribing an utterance as it was spoken the model is likely to “correct” the sentence by omitting or misclassifying tokens that may carry clinical relevance\u003csup\u003e\u003cspan additionalcitationids=\"CR17 CR18\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e–\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Failure to transcribe repeated syllables, mispronounced words, or pauses may obscure patterns crucial for detecting disease progression or symptom severity.\u003c/p\u003e \u003cp\u003eWhile recent ASR architectures, particularly large-scale, transformer-based models like OpenAI’s Whisper and Meta’s MMS have demonstrated improved robustness to diverse speech \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, their real-world performance on disordered clinical speech remains under-investigated. Evaluating these systems using conventional metrics such as Word Error Rate (WER) provides a partial view of their performance\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. WER captures surface-level discrepancies but falls short of revealing how well the ASR preserves syntactic and grammatical structure: an aspect critical in clinical Natural Language Processing (NLP) tasks such as information extraction, temporal reasoning, and diagnosis classification.\u003c/p\u003e \u003cp\u003eTo address these challenges, we conducted a multi-faceted evaluation of eight state-of-the-art commercial ASR systems on a curated dataset comprising speech samples from individuals with dementia and healthy older adults (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Our evaluation strategy integrated four complementary analyses. First, we applied Word Error Rate (WER), the standard benchmark for transcription accuracy, which measures token-level discrepancies through insertions, deletions, and substitutions. Second, we implemented Part-of-Speech (POS) distribution analysis, a syntactic-level approach that compares the frequency of grammatical categories such as nouns, verbs, and auxiliaries between ASR outputs and the ground truth. This method highlights whether the structural composition of utterances is retained, even when individual word choices differ. Third, we introduced a novel metric, POS-sqWER (Part-of-Speech sequence Word Error Rate), which blends POS tagging with the sequential alignment logic of WER. By computing the Levenshtein distance (a measure of mismatches)\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e between of POS tags from ASR and reference transcripts, POS-sqWER captures both syntactic fidelity and positional accuracy. Fourth, we implemented a stutter detection analysis to evaluate the ability of ASR models to accurately identify and reflect dysfluent speech patterns.\u003c/p\u003e \u003cp\u003e[Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e here]\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFinally, we incorporated a sentence-level embedding similarity analysis using large language models to evaluate semantic coherence between ASR outputs and the ground truth. This approach captures higher-level linguistic alignment, offering a metric sensitive to the overall meaning of an utterance even in the presence of lexical or syntactic noise. By comparing vector embeddings of sentences, we assessed the extent to which ASRs preserve potentially clinically relevant performance (e.g. word fragments, repetitions) and the original meaning of the speaker’s utterances, which are crucial for diagnosing or monitoring conditions that affect speech or language.\u003c/p\u003e \u003cp\u003eRelated Work\u003c/p\u003e \u003cp\u003eAutomatic speech recognition (ASR) is increasingly used to support clinical documentation and analysis of patient-clinician conversations, yet most evaluations often emphasize headline word error rate (WER), which can treat all errors equally and may not capture differences in clinical impact or meaning preservation\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Reviews of clinical ASR report wide WER ranges (~ 18–63%) in realistic settings\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, underscoring that a single aggregate error rate obscures what goes wrong and how it affects downstream clinical tasks.\u003c/p\u003e \u003cp\u003eTo address the limitations of WER, meaning- and task-oriented evaluations have emerged. “Semantic-WER” and related semantic distance metrics capture transcript usefulness by rewarding meaning preservation over literal n-gram overlap\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. In parallel, studies pairing ASR with medical NLP that evaluate clinical entity-level accuracy (e.g., medications, problems) have shown that two systems with similar WER can differ substantially in clinical utility, especially on accented speech and specialized terminology\u003csup\u003e\u003cspan additionalcitationids=\"CR28\" citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e–\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. Despite these advances, prior clinical ASR evaluations rarely quantify how transcripts preserve the grammatical scaffolding clinicians rely on (e.g., auxiliaries, modals, negation cues).\u003c/p\u003e \u003cp\u003eParts-of-speech in ASR research often attempts to describe error classes, not to score ASR output itself. For example, POS-aware analyses have been used to study the impact of different word classes on conversational user-modelling, but they stop short of defining POS-centric evaluation measures for ASR quality\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. We did not find clinical ASR studies that (i) measure distributional shifts in POS categories or (ii) penalize local POS-sequence disruptions as first-class metrics.\u003c/p\u003e \u003cp\u003eThis paper fills that gap by examining the accuracy of contemporary ASR systems in clinical speech cohorts by introducing two complementary metrics: POS-MDE (mean distribution error over POS categories) quantifies whether an ASR preserves the global grammatical profile of clinical discourse, and POS-sqWER (sequence-aware WER on POS tags) penalizes local syntactic mis-orderings that can flip meaning. Used alongside WER and entity-centric scores, these metrics provide a clinically relevant lens for selecting ASR systems that not only recognize words but also preserve grammatical signals crucial for interpretation, temporality, and recognition of clinically relevant communication features.\u003c/p\u003e "},{"header":"Methods","content":"\u003cp\u003e Participants and speech tasks\u003c/p\u003e\u003cp\u003eAll language samples were conducted in English, and most participants were native speakers of Australian English. Participants completed a variety of structured and semi-structured speech elicitation tasks, which differed between groups depending on the dataset source. Participants described their daily activities or personal experiences (Monologue) and completed a picture description task (The Cookie Theft Description). Detailed demographic information is provided in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. All audios were recorded in clinical or research environments at 44.1 kHz sampling rate and underwent manual quality control for completeness.\u003c/p\u003e\u003cp\u003e[Insert Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e here]\u003c/p\u003e\u003cp\u003eTable 1 Participants demographic and experiment tasks.\u003c/p\u003e\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"724\" height=\"265\"\u003e\u003c/p\u003e\u003cp\u003eThis study was approved by the Eastern Health Human Research Ethics Committee (HREC), Reference No. E21-1314-23343 (ERM ID: 23343), and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants or their legally authorised representatives prior to participation.\u003c/p\u003e\u003cp\u003eAutomatic speech recognition systems\u003c/p\u003e\u003cp\u003eTo evaluate transcription performance across various automatic speech recognition (ASR) systems, we employed eight contemporary ASR models, representing both open-source and proprietary systems (shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). These systems were selected based on their recent performance in general-purpose and clinical domains, multilingual capabilities, and model architecture diversity:\u003c/p\u003e\u003cp\u003e[Insert Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e here]\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDescription of ASR models evaluated in this study.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e \u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eASR Model\u003c/p\u003e \u003c/th\u003e\u003cth colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eAWS Transcribe\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eAmazon’s cloud-based ASR platform, evaluated in its default medical transcription settings.\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eOpenAI Whisper (large)\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eTransformer-based multilingual model trained on 680,000 hours of audio; robust to noise and accents.\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eNVIDIA Nemo\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eResearch-grade transformer ASR model with general-purpose transcription capabilities.\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eAssemblyAI\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eCommercial ASR API with support for medical domain customization and disfluency filtering.\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eMicrosoft Phi-4\u003csup\u003e34\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eInstruction-following multimodal ASR model capable of speech-to-text with advanced contextual understanding.\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eCrisperWhisper \u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e,\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eFine-tuned Whisper variant optimized for clinical and disfluent speech, designed for medical applications.\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eMeta MMS-1B-all\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eMultilingual ASR model trained on over 1,000 languages, developed by Meta (Facebook).\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eGoogle ASR\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eGoogle’s medical transcription model using the medical conversation setting for healthcare dialogue.\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003eWe used the raw ASR model transcripts in our analysis, with no modifications or corrections, to allow for accurate comparison with the manual ground truth transcriptions.\u003c/p\u003e\u003cp\u003eManual Transcription Annotation\u003c/p\u003e\u003cp\u003eThe transcription framework was based on guidelines by\u003csup\u003e39–41\u003c/sup\u003e. The framework was designed to be compatible with Universal dependencies (UD) and TalkBank\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e, who each provide tools for language sample analysis. Our framework guides users on how to represent disfluencies. For example, word fragments are represented with a hyphen (‘I w- went to the park’), and incomplete utterances are following by an ellipsis (‘why don’t you put it … don’t do that!’). These serve as critical ground truth markers for evaluating disfluency detection performance in our subsequent analyses.\u003c/p\u003e\u003cp\u003eThe manual ground truth transcriptions were produced and quality-checked by three speech pathologists with further quality checks by a clinical linguist. For a random sample of 49 of the transcripts, two transcribers independently produced a transcript by listening to the audio recording and following the transcription framework. They discussed with each other any discrepancies between their versions and resolved them through consensus to generate a finalized, accurate transcript. A further 49 were then transcribed by one transcriber by listening to the audio recording and following the transcription framework. Initial transcripts for all audio files had previously been generated by crowd-sourced workers via Amazon Mechanical Turk (MTurk); these had later been reviewed and corrected by in-house annotators. These had not followed any specific framework and typically did not include clinically relevant disfluencies like word fragments or repetitions, so they were not fit for purpose. The transcribers corrected the remaining 210 texts initially formulated by MTurk transcription by listening to the audio recording and following our transcription framework. This was designed to reduce transcription time, while maintaining high quality.\u003c/p\u003e\u003cp\u003ePreprocessing and Transcript Normalization\u003c/p\u003e\u003cp\u003eAll ASR-generated and manually transcribed texts were subjected to a standardized set of normalization and pre-processing procedures to ensure consistency and comparability across evaluation metrics. First, all text was converted to lowercase to mitigate case-based mismatches. Punctuation and special characters such as commas, periods, brackets, and other non-linguistic symbols were removed to prevent unintended lexical artifacts. Common filler words, including vocalized hesitations like “um” and “uh,” were retained to preserve disorder-specific speech characteristics in the transcription. Any extraneous whitespace introduced during transcription or formatting was trimmed. Numeric values were converted to their corresponding textual forms (e.g., “2” to “two”) to maintain alignment consistency across tools and methods. Finally, all transcripts were sorted by filenames in ascending alphanumeric to ensure consistent matching between manual and ASR transcriptions throughout the analysis pipeline.\u003c/p\u003e\u003cp\u003eEvaluation Metrics\u003c/p\u003e\u003cp\u003eTo evaluate and compare ASR performance against the manually transcribed ground truth, we employed a combination of surface-level and linguistic-level metrics, listed below and each explained further in the following sections.\u003c/p\u003e\u003cp\u003eWord Error Rate (WER)\u003c/p\u003e\u003cp\u003eWord Error Rate (WER) is a widely accepted and standardized metric\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e used to evaluate the overall transcription accuracy of Automatic Speech Recognition (ASR) systems. It computes the minimum number of insertions (I), deletions (D), and substitutions (S) required to transform the ASR-generated text into the ground truth, normalized by the total number of words in the ground truth (N). WER offers a high-level, interpretable measure of ASR performance, enabling straightforward comparisons across systems. It is easy to calculate and widely reported, making it ideal for benchmarking tasks.\u003c/p\u003e\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"277\" height=\"81\" style=\"text-align: start; color: rgb(0, 0, 0); background-color: rgb(255, 255, 255); font-size: medium; font-family: \u0026quot;\u0026quot;;\"\u003e\u003cbr\u003e\u003c/p\u003e\u003cp\u003eWER provides a quick estimate of transcription fidelity. However, it does not account for semantic or syntactic nuances. It treats all errors equally without differentiating between significant and negligible changes (e.g., omitting a repeated word vs. a replacing a noun with an unrelated verb that changes both the grammatical structure and meaning of the sentence). Moreover, WER operates at the word level and fails to capture subtler structural properties of language or sentence coherence, which are particularly crucial for clinical samples.\u003c/p\u003e\u003cp\u003ePart-of-Speech (POS) Mean Distribution Error (POS-MDE)\u003c/p\u003e\u003cp\u003eTo assess the syntactic fidelity of automatic speech recognition (ASR) outputs, we conducted a Part-of-Speech (POS) distribution analysis, which evaluates how closely the grammatical structure of ASR transcriptions aligns with that of manually transcribed ground truth texts. This analysis focuses on the frequency distribution of 13 POS categories, including nouns, verbs, and conjunctions (see Supplementary Table\u0026nbsp;1).\u003c/p\u003e\u003cp\u003eEach transcript was POS-tagged using a natural language processing library Stanza\u003csup\u003e\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e, and the proportion of each POS tag was calculated as a percentage of total tokens in the transcript. To quantify the divergence from ground truth, we computed the following difference score for each POS category:\u003c/p\u003e\u003cp\u003eDiff = ASR% − GT%\u003c/p\u003e\u003cp\u003ewhere ASR% denotes the normalized frequency of a given POS tag in the ASR output, and GT% refers to the same in the human-annotated ground truth.\u003c/p\u003e\u003cp\u003eThis method captures low-level grammatical discrepancies that may not be evident through surface-level metrics such as Word Error Rate (WER). By analyzing syntactic token distributions, we can detect whether the ASR system correctly retains functional elements like auxiliaries, determiners, and prepositions, which are often critical in clinical communication. However, while POS distribution analysis highlights structural preservation, it abstracts away semantic meaning thus, semantically divergent but syntactically similar sentences may yield indistinguishable POS profiles. Furthermore, tagging accuracy is inherently dependent on transcript quality suggesting POS distribution analysis is best used in conjunction with other metrics to form a more comprehensive evaluation of ASR performance.\u003c/p\u003e\u003cp\u003ePart-of-Speech (POS) Sequential Word Error Rate (POS-sqWER)\u003c/p\u003e\u003cp\u003eThe Part-of-Speech Sequential Word Error Rate (\u003cem\u003ePOS-sqWER\u003c/em\u003e) is a syntactically motivated evaluation metric that measures ASR fidelity by comparing the sequence of POS tags in the ASR output to the corresponding ground truth sequence. Each transcript was first POS-tagged using a standard NLP parser (Stanza) and then aligned against its reference using Levenshtein distance\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. The following equation is applied after realignment has occurred.\u003c/p\u003e\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:POSsqWER=\\frac{{S}_{\\:}+\\:D\\:+\\:I}{{N}_{POS}}\\times\\:100\\:$$\u003c/div\u003e\u003c/div\u003e\u003cp\u003eLike WER, the above formula computes the minimum number of insertions (I), deletions (D), and substitutions (S) needed to transform the ASR-derived POS tag sequence into the reference POS tag sequence after alignment and it is normalized by N\u003csub\u003ePOS,\u003c/sub\u003e the number of POS tokens in the reference sequence.\u003c/p\u003e\u003cp\u003eBy focusing on grammatical token types instead of surface words, POS-sqWER focuses on syntactic disruptions such as omission of functional words (e.g., prepositions, determiners), and reordering of syntactic units (e.g., misplaced conjunctions).\u003c/p\u003e\u003cp\u003eThis syntactic-level evaluation enables comparisons across ASR systems, especially in populations where syntactic degradation is a clinical marker (e.g., dementia, aphasia). As such, POS-sqWER complements traditional WER by offering a fine-grained lens on grammatical fidelity, supporting more informed model selection for clinical NLP and diagnostic applications.\u003c/p\u003e\u003ch3\u003eEXAMPLE:\u003c/h3\u003e\u003cp\u003eGround Truth (GT): \u003cem\u003e“The patient is experiencing frequent nausea and occasional vomiting.”\u003c/em\u003e\u003c/p\u003e\u003cp\u003eASR Output: \u003cem\u003e“The patients experiencing frequent noises and occasional bombing.”\u003c/em\u003e\u003c/p\u003e\u003cp\u003e[Insert Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e here]\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExample sentence and POS analysis.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e \u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eWord (GT)\u003c/p\u003e \u003c/th\u003e\u003cth colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003ePOS (GT)\u003c/p\u003e \u003c/th\u003e\u003cth colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003eWord (ASR)\u003c/p\u003e \u003c/th\u003e\u003cth colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003ePOS (ASR)\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eThe\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eDET\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003eThe\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003eDET\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003epatient\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eNOUN\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003epatients\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003eNOUN\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eis\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eAUX\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003e— (dropped)\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003e—\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eexperiencing\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eVERB\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003eexperiencing\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003eVERB\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003efrequent\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eADJ\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003efrequent\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003eADJ\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003enausea\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eNOUN\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003enoises\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003eNOUN\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eand\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eCCONJ\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003eand\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003eCCONJ\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003eoccasional\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eADJ\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003eoccasional\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003eADJ\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colname=\"c1\" style=\"text-align: left;\"\u003e \u003cp\u003evomiting\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c2\" style=\"text-align: left;\"\u003e \u003cp\u003eNOUN\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c3\" style=\"text-align: left;\"\u003e \u003cp\u003ebombing\u003c/p\u003e \u003c/td\u003e\u003ctd colname=\"c4\" style=\"text-align: left;\"\u003e \u003cp\u003eNOUN\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003ch2\u003eWER = (3 + 1 + 0) / 9 = 0.44\u003c/h2\u003e\u003cp\u003ePOS-sqWER = (0 + 1 + 0) / 9 = 0.11\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e illustrates an example in which WER is high while POS-sqWER remains lower, as all word substitutions occur within the same part-of-speech category as the ground truth. As a result, the grammatical structure of the ASR transcript is preserved, even though the semantic content of the sentence is substantially altered. This distinction highlights how POS-sqWER captures differences between grammatical consistency and semantic distortion that are not reflected by standard WER.\u003c/p\u003e\u003cp\u003eStutter Detection Error (SDE)\u003c/p\u003e\u003cp\u003eWe quantified stutter detection using paired ground-truth (GT) and ASR transcripts, restricting analysis to sentences that contained ≥ 1 GT stutter (sentences without GT stutters were excluded). Human annotators marked GT stutters with hyphens within or between tokens (e.g., “i-i-is”, “t- thing”, “ca- casey”), and we preserved hyphens during preprocessing. ASR outputs, by contrast, did not use hyphens; when stutters were present, they appeared as repeated tokens or character elongations (e.g., “i i i is”, “ttthing”, “c casey”). We operationalized a GT stutter event as any partial word immediately followed by a hyphen, with or without an intervening space before the completed word. For each GT event we extracted the target word (e.g., “ca- casey” → “casey”; “i-i-is” → “is”) and searched the aligned ASR sentence for a corresponding stutter realization of the same target, accepting token repeats or character elongations. To accommodate minor orthographic variation, matches were confirmed with fuzzy lexical matching (threshold = 90). We maintained an exclusion list for surface partials that occur fluently without reflecting a stutter (e.g., “we went”, “up upon”, “for forty”, “an animal”, “we were”, “not noticing”, “was washing”), which were \u003cb\u003enot\u003c/b\u003e credited as detections. When ASR produced more stutter-like forms than GT, only matches to GT targets were counted as correct (extras were ignored). For each sentence, we computed trials = number of GT stutter events, correct \u003cb\u003e=\u003c/b\u003e number matched in ASR, errors \u003cb\u003e=\u003c/b\u003e trials − correct, and defined SDE = errors / trials (Accuracy = 1 − SDE).\u003c/p\u003e\u003cp\u003eSentence Embedding Distance Analysis Using Cosine Distance\u003c/p\u003e\u003cp\u003eTo assess the semantic fidelity of ASR outputs beyond lexical or syntactic similarity, we employed a sentence embedding-based evaluation approach. This method maps both the ASR transcription and the corresponding ground truth sentences into high-dimensional semantic representations, allowing for a quantitative comparison of their meanings. In this study, we used the \"all-MiniLM-L6-v2\" model from Huggingface’s \u003cem\u003eSentenceTransformers\u003c/em\u003e library\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e,\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e, a transformer-based sentence encoder. This model is trained via knowledge distillation to approximate the performance of larger models while maintaining high efficiency, making it particularly suitable for real-time or large-scale clinical applications\u003csup\u003e\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e. Each transcript, both ground truth and ASR-generated was encoded into a 384-dimensional embedding.\u003c/p\u003e\u003cp\u003eThe cosine distance between the embeddings of each ASR sentence and its ground truth sentence was computed providing a semantic similarity score. This approach captures similarities between paraphrased or restructured content that traditional word-level metrics like WER or POS-based scores will always categorize as errors. For example, while WER penalizes synonym substitutions or reordered clauses, sentence embeddings can tolerate such variations if the overall meaning remains intact.\u003c/p\u003e\u003cp\u003eStatistical methods\u003c/p\u003e\u003cp\u003eWe analysed five metrics (WER, POS-sqWER, POS_MDE, SDE, and cosine distance) using mixed-effects models to account for repeated measures by recording. For WER and POS-sqWER, we fit linear mixed-effects models with fixed effects for ASR Model, Group (HC vs DEM), and their interaction, and a random intercept for \u003cem\u003eFilename\u003c/em\u003e. To stabilise variance and accommodate zeros, outcomes were modelled on the log1p scale and then back-transformed to percentages (i.e., \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:100\\times\\:(\\text{e}\\text{x}\\text{p}(\\widehat{\\mu\\:})-1)\\)\u003c/span\u003e\u003c/span\u003e). Extreme values were trimmed at 150% prior to modelling to reduce undue influence of outliers. We report model-estimated marginal means (“adjusted” values) and 95% CIs per ASR model × Group, with Tukey-adjusted pairwise comparisons.\u003c/p\u003e\u003cp\u003eFor POS_MDE, we fit a linear mixed-effects model with fixed effects for ASR Model, Group, POS Tag (13 levels), and all two- and three-way interactions, plus a random intercept for \u003cem\u003eFilename\u003c/em\u003e. POS_MDE was analysed on the log1p scale and back-transformed for presentation. We examined planned, Tukey-adjusted pairwise contrasts within Group × POS Tag to localise differences among ASR models.\u003c/p\u003e\u003cp\u003eFor SDE, we used a binomial GLMM on sentence-level counts (errors vs. Correct), with fixed effects for ASR Model, Group, and their interaction, and a random intercept for \u003cem\u003eFilename\u003c/em\u003e when supported by the data. Results are reported as estimated marginal means on the response scale (with 95% CIs) and as Tukey/Holm-adjusted odds-ratio contrasts among ASR models.\u003c/p\u003e\u003cp\u003eFor cosine distance, we fit a linear mixed-effects model with ASR Model, Group, and their interaction as fixed effects and a random intercept for \u003cem\u003eFilename\u003c/em\u003e. Lower values indicate better semantic fidelity; adjusted means and Tukey-adjusted pairwise tests were computed on the raw scale.\u003c/p\u003e\u003cp\u003eAcross models, we verified standard assumptions: approximate normality and homoscedasticity of residuals on the transformed scale for lmms; binomial variance and absence of complete separation for the GLMM; and conditional independence given random effects. Denominator degrees of freedom used Satterthwaite or Kenward–Roger approximations where applicable. Estimated marginal means and multiplicity-controlled comparisons were obtained from \u003cem\u003eemmeans\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e. Analyses were conducted in R (e.g., \u003cem\u003elme4/lmertest\u003c/em\u003e\u003csup\u003e48\u003c/sup\u003e, \u003cem\u003eglmmtmb\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e).\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eTo evaluate the transcription accuracy of each ASR system, we computed the WER by aligning each ASR-generated transcript with its corresponding ground truth transcript.\u003c/p\u003e \u003cp\u003e[Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e here]\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows that the model-adjusted WER had a strong main effect of ASR model \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left[F\\right(\\text{7,2106.47})=290.03,p\u0026lt;0.001]\\:\\)\u003c/span\u003e\u003c/span\u003eand Group \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left[F\\right(\\text{1,303.09})=152.53,p\u0026lt;0.001]\\)\u003c/span\u003e\u003c/span\u003e, and a significant ASR Model \u0026times; Group interaction \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left[F\\right(\\text{7,2106.47})=16.34,p\u0026lt;0.001]\\)\u003c/span\u003e\u003c/span\u003e. Given the interaction, we examined ASR differences within each group using Tukey-adjusted pairwise comparisons.\u003c/p\u003e \u003cp\u003eWithin DEM, Google had the highest adjusted WER and was significantly worse than every other system (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\u0026lt;0.001\\)\u003c/span\u003e\u003c/span\u003e); Meta and Microsoft were also significantly lower than the top performers (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\\le\\:0.001\\)\u003c/span\u003e\u003c/span\u003e). A leading cluster consisted of CrisperWhisper, AWS, AssemblyAI, and Nemo: CrisperWhisper significantly (WER=16.1) outperformed AssemblyAI and Nemo (both \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\\le\\:0.001\\)\u003c/span\u003e\u003c/span\u003e) and AWS was statistically indistinguishable from AssemblyAI and Nemo (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p=0.24\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p=0.40\\)\u003c/span\u003e\u003c/span\u003e, respectively).\u003c/p\u003e \u003cp\u003eWithin HC, Google again had the highest adjusted WER and was significantly lower than all others (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\u0026lt;0.001\\)\u003c/span\u003e\u003c/span\u003e); Meta was also significantly worse than the better systems (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\u0026lt;0.001\\)\u003c/span\u003e\u003c/span\u003e). The lowest WERs were obtained by AWS (WER=6.94) and CrisperWhisper (WER=6.76), both outperforming Microsoft, Nemo, and Whisper (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\\le\\:0.006\\)\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTaken together, these results indicate that performance rankings vary by group, with Google and Meta exhibiting significantly higher error rates, and a top tier formed by CrisperWhisper and AWS across both groups, alongside AssemblyAI and Nemo in DEM and AssemblyAI, Microsoft, Nemo and Whisper forming a statistically similar mid-tier in HC.\u003c/p\u003e \u003cp\u003e[Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e here]\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the Model-adjusted POS-sqWERs showed strong main effects of ASR Model \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left[F\\right(\\text{7,2128})=150.15,p\u0026lt;0.001]\\:\\)\u003c/span\u003e\u003c/span\u003eand Group \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left[F\\right(\\text{1,304})=153.72,p\u0026lt;0.001]\\)\u003c/span\u003e\u003c/span\u003e, and a significant ASR Model \u0026times; Group interaction \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left[F\\right(\\text{7,2128})=19.46,p\u0026lt;0.001]\\)\u003c/span\u003e\u003c/span\u003e. Given the interaction, we compared ASRs within each group using Tukey adjustment.\u003c/p\u003e \u003cp\u003eWithin DEM, Google and Microsoft had the highest POS-sqWER and were indistinguishable \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:(p=1.00)\\)\u003c/span\u003e\u003c/span\u003e, each significantly worse than AWS, CrisperWhisper, Nemo, AssemblyAI, and Whisper (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{p}\\text{}\\text{\u0026lt;}\\text{}0.001)\\)\u003c/span\u003e\u003c/span\u003e. Meta was also worse than the better-performing systems (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{p}\\text{}\\text{\u0026lt;}\\text{}0.001)\\)\u003c/span\u003e\u003c/span\u003e. A leading cluster comprised AWS and CrisperWhisper (no difference, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p=0.997\\)\u003c/span\u003e\u003c/span\u003e); Nemo did not differ significantly from AWS \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left(p=0.30\\right)\\:\\)\u003c/span\u003e\u003c/span\u003eor CrisperWhisper \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:(p=0.056)\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eWithin HC, Google again had the highest POS-sqWER, significantly worse than every other ASR (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{p\u0026lt;}0.001)\\)\u003c/span\u003e\u003c/span\u003e. Meta was also high and worse than AWS, CrisperWhisper, Microsoft, Nemo, and Whisper (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{p}\\text{}\\text{\u0026le;}\\text{}0.001)\\)\u003c/span\u003e\u003c/span\u003e. AWS and CrisperWhisper formed the top tier with no difference between them \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:(p=0.999)\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eOverall, across groups, AWS and CrisperWhisper consistently yielded the lowest adjusted POS-sqWER, as with WER. Nemo performed next best, followed by AssemblyAI and Whisper. Microsoft had middling performance with health control samples but had almost the worst performance for samples from participants with dementia. Google and Meta exhibited the highest POS-sqWER, with Google repeatedly the worst.\u003c/p\u003e \u003cp\u003e[Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e here]\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e provides a graphical overview of the POS_MDE results. There were robust main effects of ASR model, Group, and POS tag: ASR_Model, F(7, 27,243.0)\u0026thinsp;=\u0026thinsp;68.06, p\u0026thinsp;\u0026lt;\u0026thinsp;2.2\u0026times;10⁻\u0026sup1;⁶; Group, F(1, 297.4)\u0026thinsp;=\u0026thinsp;27.25, p\u0026thinsp;=\u0026thinsp;3.36\u0026times;10⁻⁷; POS_Tag, F(12, 27,251.1)\u0026thinsp;=\u0026thinsp;82.19, p\u0026thinsp;\u0026lt;\u0026thinsp;2.2\u0026times;10⁻\u0026sup1;⁶. Two-way interactions were also significant: ASR_Model \u0026times; Group, F(7, 27,243.0)\u0026thinsp;=\u0026thinsp;10.51, p\u0026thinsp;=\u0026thinsp;2.97\u0026times;10⁻\u0026sup1;\u0026sup3;; ASR_Model \u0026times; POS_Tag, F(84, 27,242.9)\u0026thinsp;=\u0026thinsp;1.60, p\u0026thinsp;=\u0026thinsp;4.44\u0026times;10⁻⁴; Group \u0026times; POS_Tag, F(12, 27,251.1)\u0026thinsp;=\u0026thinsp;23.29, p\u0026thinsp;\u0026lt;\u0026thinsp;2.2\u0026times;10⁻\u0026sup1;⁶. The three-way interaction ASR_Model \u0026times; Group \u0026times; POS_Tag was not significant, F(84, 27,242.9)\u0026thinsp;=\u0026thinsp;0.91, p\u0026thinsp;=\u0026thinsp;0.707.\u003c/p\u003e \u003cp\u003eTaken together, these effects indicate (i) overall differences in POS_MDE among ASR systems, (ii) higher POS_MDE in DEM vs HC on average, and (iii) substantial tag-specific variation, with the pattern of ASR differences varying by group and POS tag. Planned, Tukey-adjusted pairwise comparisons within Group \u0026times; POS_Tag show where models differ. Illustrative contrasts (negative estimates indicate the first model has lower POS_MDE than the second):\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eAdjectives (ADJ): differences were small; only a few contrasts reached significance (e.g., DEM: AssemblyAI\u0026thinsp;\u0026lt;\u0026thinsp;Google, estimate\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.0086, p\u0026thinsp;=\u0026thinsp;0.0045; AWS\u0026thinsp;\u0026lt;\u0026thinsp;Google, \u0026minus;\u0026thinsp;0.0077, p\u0026thinsp;=\u0026thinsp;0.0169).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAdpositions (ADP): several models exhibited lower POS_MDE than Google (e.g., DEM: Microsoft\u0026thinsp;\u0026lt;\u0026thinsp;Google, \u0026minus;\u0026thinsp;0.0096, p\u0026thinsp;=\u0026thinsp;0.0005; HC: multiple models vs Google trended lower, several p\u0026thinsp;\u0026le;\u0026thinsp;0.0065).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ePronouns (PRON) and Nouns (NOUN): consistent and sizable gaps favored multiple systems over Google in both groups (e.g., DEM PRON: AssemblyAI\u0026thinsp;\u0026lt;\u0026thinsp;Google, \u0026minus;\u0026thinsp;0.0176, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001; AWS\u0026thinsp;\u0026lt;\u0026thinsp;Google, \u0026minus;\u0026thinsp;0.0114, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001; DEM NOUN: AssemblyAI\u0026thinsp;\u0026lt;\u0026thinsp;Google, \u0026minus;\u0026thinsp;0.0224, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001; CrisperWhisper\u0026thinsp;\u0026lt;\u0026thinsp;Google, \u0026minus;\u0026thinsp;0.0188, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eVerbs (VERB): again, many models showed lower POS_MDE than Google (e.g., DEM: AssemblyAI\u0026thinsp;\u0026lt;\u0026thinsp;Google, \u0026minus;\u0026thinsp;0.0107, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001; CrisperWhisper\u0026thinsp;\u0026lt;\u0026thinsp;Google, \u0026minus;\u0026thinsp;0.0122, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eInterjections (INTJ): Meta was notably higher POS_MDE than several models in DEM (e.g., DEM: AssemblyAI\u0026thinsp;\u0026lt;\u0026thinsp;Meta, \u0026minus;\u0026thinsp;0.0144, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001; AWS\u0026thinsp;\u0026lt;\u0026thinsp;Meta, \u0026minus;\u0026thinsp;0.0189, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001), while most differences in HC were not significant.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eDeterminers (DET): Microsoft showed higher POS_MDE than many peers in DEM (e.g., DEM: AWS\u0026thinsp;\u0026lt;\u0026thinsp;Microsoft, \u0026minus;\u0026thinsp;0.0156, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001; CrisperWhisper\u0026thinsp;\u0026lt;\u0026thinsp;Microsoft, \u0026minus;\u0026thinsp;0.0166, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001), whereas differences in HC were largely nonsignificant.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eBecause lower POS_MDE indicates a closer match between ASR-predicted and human POS distributions, the above contrasts suggest (a) DEM speech increases POS distributional mismatch relative to HC; (b) some systems (e.g., Google in several tags; Microsoft for DET in DEM) exhibit larger mismatches; and (c) the ranking is tag-dependent, underscoring the value of POS-resolved evaluation.\u003c/p\u003e \u003cp\u003e[Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e here]\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSDE was modelled with a binomial GLMM including ASR Model, Group (HC vs DEM), and their interaction. There was a significant main effect of ASR Model (Type-III Wald χ\u0026sup2;(7)\u0026thinsp;=\u0026thinsp;23.76, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.0013), no main effect of Group (χ\u0026sup2;(1)\u0026thinsp;=\u0026thinsp;0.31, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.581), and no ASR\u0026times;Group interaction (χ\u0026sup2;(7)\u0026thinsp;=\u0026thinsp;4.47, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.724). Pairwise differences are reported as odds ratios (OR). For a contrast written as \u0026lsquo;Model1 / Model2,\u0026rsquo; OR\u0026thinsp;\u0026lt;\u0026thinsp;1 means Model1 has lower odds of SDE than Model2; OR\u0026thinsp;\u0026gt;\u0026thinsp;1 means Model1 has higher odds of SDE than Model2.\u003c/p\u003e \u003cp\u003eWithin DEM, several contrasts were significant after Tukey adjustment (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). Notably, CrisperWhisper (SDE\u0026thinsp;=\u0026thinsp;62.76) showed substantially lower SDE than multiple models, for example, CrisperWhisper vs Google OR\u0026thinsp;=\u0026thinsp;0.19, \u003cem\u003ep\u003c/em\u003e\u0026lt;.0001; CrisperWhisper vs Meta OR\u0026thinsp;=\u0026thinsp;0.19, \u003cem\u003ep\u003c/em\u003e\u0026lt;.0001; CrisperWhisper vs Microsoft OR\u0026thinsp;=\u0026thinsp;0.28, \u003cem\u003ep\u003c/em\u003e\u0026lt;.0001; and CrisperWhisper vs Whisper OR\u0026thinsp;=\u0026thinsp;0.11, \u003cem\u003ep\u003c/em\u003e\u0026lt;.0001. AWS also had lower SDE than Whisper (AWS vs Whisper OR\u0026thinsp;=\u0026thinsp;0.24, \u003cem\u003ep\u003c/em\u003e\u0026lt;.0001). By contrast, AssemblyAI exhibited higher SDE than several peers (e.g., AssemblyAI vs AWS OR\u0026thinsp;=\u0026thinsp;4.04, \u003cem\u003ep\u003c/em\u003e\u0026lt;.0001; AssemblyAI vs CrisperWhisper OR\u0026thinsp;=\u0026thinsp;8.49, \u003cem\u003ep\u003c/em\u003e\u0026lt;.0001; AssemblyAI vs Nemo OR\u0026thinsp;=\u0026thinsp;3.93, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.0001). Differences among Google, Meta, and Microsoft were generally not significant (e.g., Google vs Meta OR\u0026thinsp;\u0026asymp;\u0026thinsp;1.00, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.00).\u003c/p\u003e \u003cp\u003eWithin HC, no pairwise comparison reached significance after Tukey adjustment (all \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026ge;\u0026thinsp;0.466), consistent with the non-significant main effect of Group and the null ASR\u0026times;Group interaction. Taken together, SDE varied by ASR Model overall, with CrisperWhisper (and often AWS) showing lower odds of SDE than several alternatives in the DEM group, while Whisper and AssemblyAI tended to fare worse in those same comparisons. No reliable ASR ordering emerged within HC.\u003c/p\u003e \u003cp\u003e[Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e here]\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eA linear mixed-effects model (ASR_Model \u0026times; Group; random intercept for Filename) showed strong main effects of ASR model and Group, and a significant interaction (ASR_Model: \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:F\\left(\\text{7,2128}\\right)=300.31,p\u0026lt;2.2\\times\\:{10}^{-16}\\)\u003c/span\u003e\u003c/span\u003e; Group: \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:F\\left(\\text{1,323.4}\\right)=60.31,p=1.08\\times\\:{10}^{-13}\\)\u003c/span\u003e\u003c/span\u003e; ASR_Model\u0026times;Group: \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:F\\left(\\text{7,2128}\\right)=11.23,p=4.93\\times\\:{10}^{-14}\\)\u003c/span\u003e\u003c/span\u003e). Within-group Tukey comparisons (on the raw cosine-distance scale; lower is better) indicated that Google had substantially higher (worse) cosine distances than every other ASR in both groups (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\u0026lt;0.0001;\\text{F}\\text{i}\\text{g}.\\:6\\)\u003c/span\u003e\u003c/span\u003e). In DEM, AssemblyAI, CrisperWhisper, Whisper, and (to a lesser extent) AWS formed the leading tier: AssemblyAI outperformed AWS and Microsoft (both \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\\le\\:0.0012\\)\u003c/span\u003e\u003c/span\u003e) and was statistically indistinguishable from CrisperWhisper and Whisper (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\\ge\\:0.93\\)\u003c/span\u003e\u003c/span\u003e); CrisperWhisper and Whisper were also indistinguishable (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p=1.00\\)\u003c/span\u003e\u003c/span\u003e) and each beat AWS (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\\le\\:0.035\\)\u003c/span\u003e\u003c/span\u003e). Meta and Microsoft were reliably worse than this leading tier but still markedly better than Google (all contrasts vs Google \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\u0026lt;0.0001\\)\u003c/span\u003e\u003c/span\u003e). Nemo sat mid-pack\u0026mdash;better than Google ( \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\u0026lt;0.0001\\)\u003c/span\u003e\u003c/span\u003e) but worse than AssemblyAI ( \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p=0.033\\)\u003c/span\u003e\u003c/span\u003e). In HC, the same pattern held: Google was again worse than all peers (all \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\u0026lt;0.0001\\)\u003c/span\u003e\u003c/span\u003e); CrisperWhisper and Whisper were statistically similar (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p=1.00\\)\u003c/span\u003e\u003c/span\u003e), AWS was similar to Whisper (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p=1.00\\)\u003c/span\u003e\u003c/span\u003e), and Microsoft and Nemo were indistinguishable (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p=1.00\\)\u003c/span\u003e\u003c/span\u003e). Collectively, these results confirm meaningful model-wise differences in semantic fidelity (cosine distance) that vary by Group, with Google consistently underperforming and a top cluster comprising Whisper, CrisperWhisper, AssemblyAI (and often AWS) exhibiting the lowest distances.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur multi-metric evaluation highlights that model choice and model robustness depend strongly on both the linguistic target and speaker group. Across global accuracy measures, WER and POS-sqWER produced similar trends: performance was markedly better for HC than DEM, and ASR models differed substantially. In both groups Google exhibited consistently higher error rates, while AWS and CrisperWhisper formed a reliable top tier, joined by AssemblyAI and Nemo in DEM and by several mid-tier contenders (AssemblyAI, Microsoft, Nemo and Whisper) in HC. This HC\u0026ndash;DEM gap, replicated again in the cosine-distance analysis of sentence embeddings, underscores a persistent vulnerability of current systems to atypical prosody and articulation an effect likely rooted in training corpora skewed toward normative speech.\u003c/p\u003e \u003cp\u003eBeyond token-level accuracy, the POS_MDE analysis shows why \u0026ldquo;which model is best\u0026rdquo; is intrinsically task-dependent. There were strong main effects of ASR, Group, and POS tag, and the pattern of differences shifted across parts of speech: several models showed large mismatches for specific categories in DEM (e.g., Google across multiple tags; Microsoft for determiners), whereas other categories showed only small, scattered differences. Nemo was frequently among the strongest at preserving POS distributions, and Whisper and AssemblyAI were broadly competitive except for isolated tags (e.g., VERB for Whisper, SCONJ for AssemblyAI in DEM). These tag-resolved effects argue for selecting models with an eye to the linguistic constructs most critical for a given downstream analysis rather than relying on a single omnibus score.\u003c/p\u003e \u003cp\u003eStutter detection exposed a different facet of system behaviour. Although there was a significant overall ASR effect, there was no group effect and no ASR\u0026times;Group interaction, and absolute accuracies were low for all models. Within DEM, CrisperWhisper exhibited reliably lower odds of stutter-detection error than several competitors; in HC, pairwise differences were not reliable, consistent with the very small number of stutter events in that group. The uniformly low performance here indicates that disfluency modelling remains an unmet need: even systems that are accurate on words and semantics struggle to capture clinically salient disfluencies that often drive assessment decisions.\u003c/p\u003e \u003cp\u003eSemantic fidelity (cosine distance) reinforced the above conclusions while adding nuance. Again, HC transcripts were closer to the ground truth than DEM, and there were pronounced model effects with a significant interaction by Group. Google had the largest (worst) cosine distances in both groups by a wide margin. A leading cluster Whisper, CrisperWhisper, and AssemblyAI, often joined by AWS showed the smallest distances, with Meta and Microsoft clearly behind that cluster but still substantially better than Google. Nemo tended to sit mid-pack on this measure despite its strong POS behaviour. Together with WER and POS-sqWER, these findings suggest that models can trade off token-level accuracy and sentence-level semantic alignment in subtle ways; a model that is mid-tier on WER can still preserve meanings well, and vice-versa.\u003c/p\u003e \u003cp\u003ePractically, three themes emerge. First, group matters: every metric showed a degradation for DEM, signaling that off-the-shelf models need adaptation or fine-tuning to handle atypical speech. Second, metric choice matters: AWS and CrisperWhisper are consistently strong on WER and POS-sqWER; Whisper, CrisperWhisper, and AssemblyAI often lead on semantic alignment. Third, use-case matters: if the endpoint depends on specific grammatical categories (e.g., function-word distributions), Nemo or Whisper and AssemblyAI may be preferable; for general clinical transcription where token accuracy dominates, AWS and CrisperWhisper are safe defaults; for workflows requiring detection of disfluency, none of the models yet meet a clinically useful bar, with CrisperWhisper the current but limited front-runner in DEM.\u003c/p\u003e \u003cp\u003eOur study\u0026rsquo;s strength is its unified framework spanning token errors, syntactic structure, semantics, and disfluency, revealing that \u0026ldquo;accuracy\u0026rdquo; is multi-dimensional and that rankings shift when the lens changes. Limitations include relatively sparse stutter events in HC and the possibility that some POS-resolved differences reflect sampling variability at the tag level; larger, more diverse clinical corpora and model adaptation to pathological speech should narrow uncertainty bands and may alter rankings. Still, the overall pattern is clear: modern ASR systems can perform well on normative speech yet manifest predictable failure modes in clinical contexts. Model selection should therefore be purpose-built, and future development should prioritize robustness to atypical speech and explicit modelling of disfluencies if ASR is to support reliable, fair, and clinically useful inference.\u003c/p\u003e \u003cp\u003eLimitations and future research\u003c/p\u003e \u003cp\u003eOur dataset includes linguistic samples from a group of Australian healthy older adults and adults with dementia. Thus, we cannot generalize our findings to broader populations, such as children or younger adults, speakers of other languages, or other clinical conditions such as Parkinson\u0026rsquo;s disease, ALS, or developmental speech impairments. Notably, the analysis of stutter detection was constrained by the markedly low incidence of stutters in the healthy group (only 11 instances), which warrants exploration of these ASR models using language data from a group of speakers who experience stuttering or other disfluencies to evaluate their performance.\u003c/p\u003e \u003cp\u003eAlthough our evaluation spans eight ASR systems covering commercial, open-source, and fine-tuned models, this set is not exhaustive, and the inclusion of proprietary or next-generation models might yield different outcomes.\u003c/p\u003e \u003cp\u003eSecond, the syntactic and semantic analyses employed rely on general-purpose NLP models (in our case, Stanza) for POS tagging and all-MiniLM-L6-v2 for sentence embeddings. These models are not specifically trained on clinical speech and may inadequately capture the nuances of pathological language, such as irregular grammar, disfluencies, or atypical prosody. As a result, transcription errors with clinical significance, particularly those involving disfluency or language impairment, may be underestimated or mis-characterised.\u003c/p\u003e \u003cp\u003eThird, while POS-sqWER introduces valuable grammatical sensitivity into ASR evaluation, it currently assigns equal weight to all part-of-speech mismatches. This uniform weighting overlooks the differing functional and diagnostic significance of various POS categories. For example, errors in nouns or verbs may carry more clinical consequence than those involving function words like determiners or auxiliary verbs. Similarly, sentence embedding-based semantic metrics, though effective for capturing general meaning preservation, may miss fine-grained deviations that are critical in clinical interpretation, such as missed negations, disfluent insertions, or altered temporal references.\u003c/p\u003e \u003cp\u003eOur results demonstrate areas of future research in the improvement in both ASR systems and their evaluation. Expanding training datasets for ASR models to include a more balanced representation of speech, language and neurological conditions could improve available ASR models. In terms of testing and refining these models, incorporating a wide suite of performance metrics (as we have done here) could significantly enhance the diagnostic relevance and precision of ASR technologies in uses involving clinical populations.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eAcross metrics and groups, AWS and CrisperWhisper were the most consistently strong performers, while Google and Meta underperformed. AssemblyAI, Nemo, Microsoft, and Whisper generally occupied a mid-tier with performance varying by metric and, in some cases, by linguistic category. Participants with dementia were transcribed less accurately than healthy controls across nearly all measures, reinforcing the clinical gap that remains for atypical speech. POS-MDE highlighted tag-specific differences that can inform model choice when particular parts of speech matter for a given application. Stutter detection was weak for all systems, although CrisperWhisper was comparatively better in the dementia group and absolute accuracies were still low. Taken together, a multi-metric evaluation provides a more faithful and actionable picture of clinical ASR performance than reliance on WER alone.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eDeclarations of interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDrs Kathiresan and Profs Vogel, and Ms. Gasparini are employees of Redenlab Ltd., a speech neuroscience company. Ms. Zona Ling was an intern at Redenlab during this work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe acknowledge funding support through the CUREator+ Dementia and Cognitive Decline BioMedTech Incubator program supported by ANDHealth and partners, which contributed to the digital health research environment underpinning this work.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTK: Conceptualization, Data curation, Formal analysis, Methodology, Project administration, Visualization, Writing - original draft, Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eZL: Data curation, Methodology, Software, Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eLG: Data curation, Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eSEP: Data curation, Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eCT: Data curation, Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eHE: Data curation, Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eAB:\u0026nbsp;Methodology, Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eAPV: Conceptualization, Funding acquisition, Methodology, Supervision, Writing - review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData sharing statement:\u0026nbsp;\u003c/strong\u003eAll data, code, and results are available at:https://github.com/kathiresant-redenlab/ASR_MultiMetric_Evaluation\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration\u003c/strong\u003e\u003cstrong\u003eof generative AI and AI-assisted technologies in the manuscript preparation process\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this work, authors used Large Language Models (ChatGPT) to edit their writing for grammatical correctness and clarity. After using these tools, the authors reviewed and edited the content and take full responsibility for the content of the published article.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCornell, S. \u003cem\u003eet al.\u003c/em\u003e The CHiME-7 DASR Challenge: Distant Meeting Transcription with Multiple Devices in Diverse Scenarios. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.2306.13734\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.2306.13734\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShah, M. A., Noguero, D. S., Heikkila, M. A., Raj, B. \u0026amp; Kourtellis, N. Speech Robust Bench: A Robustness Benchmark For Speech Recognition. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2403.07937\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2403.07937\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJafari, Z., Andrew, M. K. \u0026amp; Rockwood, K. J. Diagnostic utility of speech-based biomarkers in mild cognitive impairment: a systematic review and meta-analysis. \u003cem\u003eAge Ageing\u003c/em\u003e 54, afaf316 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFalcetta, F. S., De Almeida, F. K., Lemos, J. C. S., Goldim, J. R. \u0026amp; Da Costa, C. A. Automatic documentation of professional health interactions: A systematic review. \u003cem\u003eArtif. Intell. Med.\u003c/em\u003e 137, 102487 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin, H. \u003cem\u003eet al.\u003c/em\u003e Identification of digital voice biomarkers for cognitive health. \u003cem\u003eExplor. Med.\u003c/em\u003e 1, 406\u0026ndash;417 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHecker, P., Steckhan, N., Eyben, F., Schuller, B. W. \u0026amp; Arnrich, B. Voice Analysis for Neurological Disorder Recognition\u0026ndash;A Systematic Review and Perspective on Emerging Trends. \u003cem\u003eFront. Digit. Health\u003c/em\u003e 4, 842301 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGarc\u0026iacute;a, A. M., De Leon, J., Tee, B. L., Blasi, D. E. \u0026amp; Gorno-Tempini, M. L. Speech and language markers of neurodegeneration: a call for global equity. \u003cem\u003eBrain\u003c/em\u003e 146, 4870\u0026ndash;4879 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao, F., Vogel, A. P., Gharahkhani, P. \u0026amp; Renteria, M. E. Speech and language biomarkers for Parkinson\u0026rsquo;s disease prediction, early diagnosis and progression. \u003cem\u003eNpj Park. Dis.\u003c/em\u003e 11, 57 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChan, J. C. S., Stout, J. C. \u0026amp; Vogel, A. P. Speech in prodromal and symptomatic Huntington\u0026rsquo;s disease as a model of measuring onset and progression in dominantly inherited neurodegenerative diseases. \u003cem\u003eNeurosci. Biobehav. Rev.\u003c/em\u003e 107, 450\u0026ndash;460 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNoffs, G. \u003cem\u003eet al.\u003c/em\u003e What speech can tell us: A systematic review of dysarthria characteristics in Multiple Sclerosis. \u003cem\u003eAutoimmun. Rev.\u003c/em\u003e 17, 1202\u0026ndash;1209 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGreen, J. R. \u003cem\u003eet al.\u003c/em\u003e Bulbar and speech motor assessment in ALS: Challenges and future directions. \u003cem\u003eAmyotroph. Lateral Scler. Front. Degener.\u003c/em\u003e 14, 494\u0026ndash;500 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePoole, M. L., Brodtmann, A., Darby, D. \u0026amp; Vogel, A. P. Motor Speech Phenotypes of Frontotemporal Dementia, Primary Progressive Aphasia, and Progressive Apraxia of Speech. \u003cem\u003eJ. Speech Lang. Hear. Res.\u003c/em\u003e 60, 897\u0026ndash;911 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuffy, J. R., Utianski, R. L. \u0026amp; Josephs, K. A. Primary progressive apraxia of speech: from recognition to diagnosis and care. \u003cem\u003eAphasiology\u003c/em\u003e 35, 560\u0026ndash;591 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCroot, K., Ballard, K., Leyton, C. E. \u0026amp; Hodges, J. R. Apraxia of Speech and Phonological Errors in the Diagnosis of Nonfluent/Agrammatic and Logopenic Variants of Primary Progressive Aphasia. \u003cem\u003eJ. Speech Lang. Hear. Res.\u003c/em\u003e 55, (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeung, W.-Z., Cross, M., Ragni, A. \u0026amp; Goetze, S. Training Data Augmentation for Dysarthric Automatic Speech Recognition by Text-to-Dysarthric-Speech Synthesis. in \u003cem\u003eInterspeech 2024\u003c/em\u003e 2494\u0026ndash;2498 (ISCA, 2024). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21437/Interspeech.2024-1645\u003c/span\u003e\u003cspan address=\"10.21437/Interspeech.2024-1645\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou, L. \u003cem\u003eet al.\u003c/em\u003e Analysis of Errors in Dictated Clinical Documents Assisted by Speech Recognition Software and Professional Transcriptionists. \u003cem\u003eJAMA Netw. Open\u003c/em\u003e 1, e180530 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Buchem, M. M. \u003cem\u003eet al.\u003c/em\u003e The digital scribe in clinical practice: a scoping review and research agenda. \u003cem\u003eNpj Digit. Med.\u003c/em\u003e 4, 57 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeleki, M., Dong, X., Kim, S. \u0026amp; Caverlee, J. Comparing ASR Systems in the Context of Speech Disfluencies. in \u003cem\u003eInterspeech 2024\u003c/em\u003e 4548\u0026ndash;4552 (ISCA, 2024). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21437/Interspeech.2024-1270\u003c/span\u003e\u003cspan address=\"10.21437/Interspeech.2024-1270\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMichot, J. \u003cem\u003eet al.\u003c/em\u003e Error-preserving Automatic Speech Recognition of Young English Learners\u0026rsquo; Language. in \u003cem\u003eProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)\u003c/em\u003e 6444\u0026ndash;6454 (Association for Computational Linguistics, Bangkok, Thailand, 2024). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2024.acl-long.348\u003c/span\u003e\u003cspan address=\"10.18653/v1/2024.acl-long.348\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRadford, A. \u003cem\u003eet al.\u003c/em\u003e Robust Speech Recognition via Large-Scale Weak Supervision. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.2212.04356\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.2212.04356\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePratap, V. \u003cem\u003eet al.\u003c/em\u003e Scaling Speech Technology to 1,000\u0026thinsp;+\u0026thinsp;Languages. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.2305.13516\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.2305.13516\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNavarro, G. A guided tour to approximate string matching. \u003cem\u003eACM Comput. Surv.\u003c/em\u003e 33, 31\u0026ndash;88 (2001).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTran, B. D., Mangu, R., Tai-Seale, M., Lafata, J. E. \u0026amp; Zheng, K. Automatic speech recognition performance for digital scribes: a performance comparison between general-purpose and specialized models tuned for patient-clinician conversations. \u003cem\u003eAMIA Annu. Symp. Proc. AMIA Symp.\u003c/em\u003e 2022, 1072\u0026ndash;1080 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMirheidari, B., Blackburn, D., Walker, T., Reuber, M. \u0026amp; Christensen, H. Dementia detection using automatic analysis of conversations. \u003cem\u003eComput. Speech Lang.\u003c/em\u003e 53, 65\u0026ndash;79 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMiner, A. S. \u003cem\u003eet al.\u003c/em\u003e Assessing the accuracy of automatic speech recognition for psychotherapy. \u003cem\u003eNpj Digit. Med.\u003c/em\u003e 3, 82 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchultz, B. G. \u003cem\u003eet al.\u003c/em\u003e Automatic speech recognition in neurodegenerative disease. \u003cem\u003eInt. J. Speech Technol.\u003c/em\u003e 24, 771\u0026ndash;779 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAfonja, T. \u003cem\u003eet al.\u003c/em\u003e Performant ASR Models for Medical Entities in Accented Speech. in \u003cem\u003eInterspeech 2024\u003c/em\u003e 2315\u0026ndash;2319 (ISCA, 2024). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21437/Interspeech.2024-2261\u003c/span\u003e\u003cspan address=\"10.21437/Interspeech.2024-2261\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLe-Duc, K. \u003cem\u003eet al.\u003c/em\u003e Medical Spoken Named Entity Recognition. in \u003cem\u003eProceedings of the\u003c/em\u003e 2025 \u003cem\u003eConference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track)\u003c/em\u003e 724\u0026ndash;783 (Association for Computational Linguistics, Albuquerque, New Mexico, 2025). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2025.naacl-industry.59\u003c/span\u003e\u003cspan address=\"10.18653/v1/2025.naacl-industry.59\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShapira, O., Chazan, S. \u0026amp; Cohen, A. D. N. Measuring the Effect of Transcription Noise on Downstream Language Understanding Tasks. in \u003cem\u003eProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)\u003c/em\u003e 29978\u0026ndash;30004 (Association for Computational Linguistics, Vienna, Austria, 2025). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2025.acl-long.1449\u003c/span\u003e\u003cspan address=\"10.18653/v1/2025.acl-long.1449\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao, J. \u003cem\u003eet al.\u003c/em\u003e A Comparative Analysis of Automatic Speech Recognition Errors in Small Group Classroom Discourse. in \u003cem\u003eProceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization\u003c/em\u003e 250\u0026ndash;262 (ACM, Limassol Cyprus, 2023). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3565472.3595606\u003c/span\u003e\u003cspan address=\"10.1145/3565472.3595606\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmazon Web Services, Inc. Amazon Transcribe. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://aws.amazon.com/transcribe\u003c/span\u003e\u003cspan address=\"https://aws.amazon.com/transcribe\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHugging Face. nvidia/stt_en_conformer_ctc_large. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://huggingface.co/nvidia/stt_en_conformer_ctc_large\u003c/span\u003e\u003cspan address=\"https://huggingface.co/nvidia/stt_en_conformer_ctc_large\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAssemblyAI. \u003cem\u003eAssemblyAI API Documentation\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.assemblyai.com/docs\u003c/span\u003e\u003cspan address=\"https://www.assemblyai.com/docs\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMicrosoft. microsoft/Phi-4-multimodal-instruct. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://huggingface.co/microsoft/Phi-4-multimodal-instruct\u003c/span\u003e\u003cspan address=\"https://huggingface.co/microsoft/Phi-4-multimodal-instruct\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWagner, L., Thallinger, B. \u0026amp; Zusag, M. CrisperWhisper: Accurate Timestamps on Verbatim Speech Transcriptions. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2408.16589\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2408.16589\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNyra Health. CrisperWhisper: A custom automatic speech recognition model based on Whisper. GitHub repository (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/nyrahealth/CrisperWhisper\u003c/span\u003e\u003cspan address=\"https://github.com/nyrahealth/CrisperWhisper\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePratap, V. et al. Scaling Speech Technology to 1,000\u0026thinsp;+\u0026thinsp;Languages. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2305.13516\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2305.13516\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoogle Cloud. \u003cem\u003eRecognize Speech by Using Medical Models (Speech-to-Text V1)\u003c/em\u003e. (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLanguage Development Project Team. 4. Utterance boundaries. (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKahane, S., Caron, B., Strickland, E. \u0026amp; Gerdes, K. Annotation guidelines of UD and SUD treebanks for spoken corpora: A proposal. in \u003cem\u003eProceedings of the 20th International Workshop on Treebanks and Linguistic Theories (TLT, SyntaxFest 2021)\u003c/em\u003e (eds Dakota, D., Evang, K. \u0026amp; K\u0026uuml;bler, S.) 35\u0026ndash;47 (Association for Computational Linguistics, Sofia, Bulgaria, 2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacWhinney, B. CHAT Manual. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.21415/3MHN-0Z89\u003c/span\u003e\u003cspan address=\"10.21415/3MHN-0Z89\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacWhinney, B. \u0026amp; Fromm, D. Language Sample Analysis With TalkBank: An Update and Review. \u003cem\u003eFront. Commun.\u003c/em\u003e 7, 865498 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNational Institute of Standards and Technology. Speech Recognition Scoring Toolkit (SCTK) and \\textttsclite. (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQi, P., Zhang, Y., Zhang, Y., Bolton, J. \u0026amp; Manning, C. D. Stanza: A Python Natural Language Processing Toolkit for Many Human Languages. in \u003cem\u003eProceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations\u003c/em\u003e 101\u0026ndash;108 (Association for Computational Linguistics, Online, 2020). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2020.acl-demos.14\u003c/span\u003e\u003cspan address=\"10.18653/v1/2020.acl-demos.14\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReimers, N. \u0026amp; Gurevych, I. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. in \u003cem\u003eProceedings of the\u003c/em\u003e 2019 \u003cem\u003eConference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)\u003c/em\u003e 3982\u0026ndash;3992 (2019). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/D19-1410\u003c/span\u003e\u003cspan address=\"10.18653/v1/D19-1410\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReimers, N. \u0026amp; Gurevych, I. Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation. in \u003cem\u003eProceedings of the\u003c/em\u003e 2020 \u003cem\u003eConference on Empirical Methods in Natural Language Processing (EMNLP)\u003c/em\u003e 4512\u0026ndash;4525 (Association for Computational Linguistics, Online, 2020). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2020.emnlp-main.365\u003c/span\u003e\u003cspan address=\"10.18653/v1/2020.emnlp-main.365\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSearle, S. R., Speed, F. M. \u0026amp; Milliken, G. A. Population Marginal Means in the Linear Model: An Alternative to Least Squares Means. \u003cem\u003eAm. Stat.\u003c/em\u003e 34, 216\u0026ndash;221 (1980).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLenth, R. V. Least-Squares Means: The \u003cem\u003eR\u003c/em\u003e Package lsmeans. \u003cem\u003eJ. Stat. Softw.\u003c/em\u003e 69, (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrooks, M., E. \u003cem\u003eet al.\u003c/em\u003e glmmTMB Balances Speed and Flexibility Among Packages for Zero-inflated Generalized Linear Mixed Modeling. \u003cem\u003eR J.\u003c/em\u003e 9, 3\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":true,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Automatic speech recognition, Natural language processing, Speech, Language, Dementia, Digital biomarkers, Clinical research","lastPublishedDoi":"10.21203/rs.3.rs-8824529/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8824529/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAutomatic speech recognition (ASR) is increasingly used in digital health, yet its reliability for populations with atypical speech, such as people with dementia, is not well characterised beyond word error rate (WER). We evaluated eight ASR systems on speech from adults with dementia and healthy older adults using WER, part-of-speech sequence agreement (POS-sqWER), part-of-speech distribution mismatch (POS-MDE), sentence-embedding cosine distance, and stutter detection error (SDE). Mixed-effects models with Tukey-corrected contrasts were used for comparison. Performance was consistently poorer for dementia speech across all metrics. AWS and CrisperWhisper showed relatively strong lexical, semantic, and syntactic fidelity, whereas Google and Meta exhibited lower accuracy. Other systems showed intermediate performance with rankings varying by metric. POS-MDE revealed syntactic distortions not captured by WER or POS-sqWER. SDE performance was low across systems. Multidimensional evaluation reveals clinically relevant linguistic distortions obscured by WER alone, supporting the need for population- and task-specific ASR validation in digital health research.\u003c/p\u003e","manuscriptTitle":"Beyond word error rate: Multidimensional evaluation of ASR performance for digital speech biomarkers","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-17 06:43:35","doi":"10.21203/rs.3.rs-8824529/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"fdf6f42c-5998-479f-9e14-092c432f4fa4","owner":[],"postedDate":"February 17th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":62942002,"name":"Health sciences/Biomarkers"},{"id":62942003,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":62942004,"name":"Health sciences/Health care"},{"id":62942005,"name":"Health sciences/Medical research"},{"id":62942006,"name":"Biological sciences/Neuroscience"}],"tags":[],"updatedAt":"2026-03-03T04:09:31+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-17 06:43:35","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8824529","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8824529","identity":"rs-8824529","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}