A Noise-Robust End-to-End Framework for Amharic Speech Recognition

A Noise-Robust End-to-End Framework for Amharic Speech Recognition | 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 Research Article A Noise-Robust End-to-End Framework for Amharic Speech Recognition Yohannes Ayana Ejigu, Tesfa Tegegne Asfaw, Surafel Amsalu Tadesse This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6419413/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 24 Dec, 2025 Read the published version in International Journal of Speech Technology → Version 1 posted 7 You are reading this latest preprint version Abstract End-to-end automatic speech recognition (ASR) offers a streamlined alternative to traditional systems that rely on multiple, separately trained language, acoustic, and pronunciation models. In this paper, we present a noise-robust, end-to-end ASR framework tailored specifically to the Amharic language. Our approach integrates a convolutional neural network (CNN), a recurrent neural network (RNN), and Connectionist Temporal Classification (CTC) to directly transcribe speech into text—bypassing the need for labor-intensive dictionary creation. We evaluate our method on a large corpus of 20,000 noisy Amharic utterances, achieving a word error rate (WER) of just 7%. This result underlines the effectiveness of our system in handling challenging acoustic conditions. By reducing complexity and manual overhead, our end-to-end model offers a practical and accurate solution for real-world deployments, with broader implications for developing ASR in other low-resource and noise-prone environments. Amharic Speech Recognition End-to-End Deep Learning Noise-Robust ASR BiGRU Connectionist Temporal Classification Real-World Noisy Environments Automatic Feature Extraction Low-Resource Languages Spectral Subtraction Subspace Filtering Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Speech recognition—often referred to as automatic speech recognition (ASR), computer speech recognition, or speech-to-text—facilitates the conversion of spoken language into written text. While it is sometimes conflated with voice recognition, which identifies the speaker rather than the words themselves, ASR focuses exclusively on the transcription of spoken utterances. Recent advancements in deep learning have prompted a surge of applications across multiple sectors, from healthcare and finance to public safety, providing immense time-saving and life-saving potential (Hanan Aldarmaki, Asad Ullah, & Nazar Zaki, 2021). Over the years, a variety of methods have been explored for constructing ASR systems, including Dynamic Time Warping (DTW), Hidden Markov Models (HMMs), Dynamic Bayesian Networks (DBNs), and artificial neural networks (ANNs) (Hebash H.O. Nasereddin, January 2018). In particular, HMMs have historically demonstrated effective performance in decoding acoustic signals into phonetic states (Kebebew, 2010). However, these approaches often rely on multiple, separately trained modules—such as pronunciation dictionaries, acoustic models, and language models—that must work in tandem. This not only increases computational and engineering complexity but can also introduce inconsistencies between frame-level training objectives and sequence-level performance (Bourlard & Morgan, 1993; Hinton et al., 2012). For instance, improvements in per-frame accuracy sometimes fail to boost overall transcription accuracy, highlighting a fundamental mismatch in training and inference objectives (Bahl et al., 1986; Jaitly et al., 2012). Furthermore, traditional pipeline systems typically require meticulous human curation of phonemic dictionaries to map words into subword units, introducing significant time and resource overhead (Graves, A., & Jaitly, N., 2014). For languages such as Amharic—where large-scale lexical databases may not be readily available or are resource-intensive to build—these dependencies present substantial barriers. Conventional speech recognition systems for Amharic often ignore noisy real-world conditions and focus on relatively clean datasets, thereby limiting their applicability (Baye, A., Tachbelie, Y., & Besacier, L., 2021). To address these challenges, we propose an end-to-end ASR framework that incorporates a recurrent neural network (RNN) as the principal component, streamlining the entire speech-recognition pipeline. By replacing separate acoustic and pronunciation modules with a single neural sequence model, our approach bypasses the need for extensive dictionary construction and aligns model training directly with the sequence-level transcription goal. Although direct waveform-to-text transcription has been explored (Graves, 2012; Jaitly & Hinton, 2011), computational overhead remains a concern, making spectrogram-based preprocessing a practical compromise. Additionally, to handle dynamic noise environments, we incorporate robust feature extraction and noise-mitigation strategies into our end-to-end pipeline. Given that current end-to-end architectures for Amharic often overlook noisy data and rarely employ deep neural methods for feature extraction (Baye et al., 2021), this research fills a crucial gap by providing a noise-aware, end-to-end ASR model for Amharic. In contrast to conventional HMM-based solutions, our approach eliminates multiple sub-pipelines—such as manual pronunciation dictionaries—thus reducing the reliance on domain experts. We also leverage Connectionist Temporal Classification (CTC) to address alignment issues that typically plague conventional systems. 1.1 Research Questions This study addresses the following key questions: RQ1: What impact does applying erosion to spectrogram features have on the accuracy of an Amharic speech recognition model? RQ2: Which noise reduction approach—spectral subtraction, subspace filtering, or a combined strategy—proves most effective for end-to-end Amharic ASR? RQ3: Between BiLSTM and BiGRU architectures, which yields superior performance and efficiency for Amharic ASR under noisy conditions? 2. Data We conducted audio recordings of individuals delivering speeches in the Sidama region, where Amharic is not their primary language. The noisy dataset consists of 400 distinct sentences, each sentence being read by 50 different individuals. In total, 20,000 sentences are used for the recordings. The noisy audios are read speech data which last from a minimum of 4 seconds to a maximum of 20 seconds. The total noisy data is around 44 hours and 46 minutes. The dataset referenced is currently accessible online under our authorship. It is readily available via the provided link below: https://drive.google.com/drive/folders/155kEo_yyJRIOyBTv8ruhHmHdNKQPx_My . For the speech recording (noisy data), a corresponding transcription was created. This process involved carefully listening to the audio and converting it into written text. The text data serves as the ground truth for the deep learning model, allowing it to learn the relationship between the audio features and the corresponding text. Capturing speech data in environments characterized by high levels of noise demands meticulous attention to hardware, software, and environmental variables to uphold the integrity and caliber of the dataset. In pursuit of accurate speech capture within such challenging conditions, we deliberately selected robust microphones engineered to excel in noise-laden settings, alongside the latest smartphone technology to enhance our recording capabilities. Our recording protocol was underpinned by the utilization of industry-standard software, affording us the capacity to faithfully capture the vocalizations of 50 distinct individuals. These software tools facilitated precise and comprehensive audio recording, ensuring the fidelity and authenticity of the dataset. Recording sessions were deliberately orchestrated across a spectrum of noisy environments, mirroring real-world scenarios with precision. These environments encompassed bustling urban locales, resonating with the clamor of vehicular traffic and bustling crowds, as well as tranquil public spaces such as parks, where the ambient soundscape embraced the melodic trills of avian inhabitants and the rustling of foliage. Additionally, recordings were made in various other noisy settings, including environments characterized by human activity at a distance. The decision to collect the corpus in the Sidama region is rooted in the area's significant linguistic and demographic characteristics. Recent studies (Shaik Johny Basha et al., 2023) report over 60,000 Amharic speakers in Ethiopia. Moreover, the Sidama region attracts individuals from diverse geographic backgrounds, fostering a rich linguistic landscape characterized by various dialects and linguistic influences. While we acknowledge the importance of capturing speech data from regions where Amharic is the primary language, we also recognize the necessity of ensuring geographic diversity within our dataset. Incorporating speakers from areas where Amharic serves as a second language is paramount for enhancing the robustness and diversity of our model. By including individuals from diverse linguistic backgrounds, we aim to improve the model's adaptability to different speech patterns and linguistic contexts, thereby strengthening its performance and usability in real-world applications. The text corpus comprises 44,937 unique words within the noisy Amharic language dataset. Additionally, the corpus encompasses more than 250 characters. To ensure linguistic balance among phonemes, a comprehensive frequency analysis of both words and characters has been conducted using a Python script. This analysis not only quantifies the occurrence of individual phonemes but also provides insights into the distribution and representation of linguistic elements within the corpus. The noisy environment in which the data was collected was deliberately designed to incorporate diverse sources of background noise. These included ambient sounds such as bird calls from a nearby park, traffic noise from passing cars, wind disturbances, and chatter from other individuals, such as conversations in a cafeteria setting. For this study, the demographic composition of the speakers primarily comprises young adults and middle-aged adults, with an equal representation of both male and female participants. It's important to note that our current research focus has led us to include a specific age range, spanning from 18 to 50 years, ensuring a balanced distribution across genders. The dataset was divided into three subsets for training, validation, and testing, with a split ratio of 80%, 10%, and 10%, respectively. The transcribed text paired with the file name of the audio is presented in Fig. 1 below. The collected data is analyzed to gain insight into speech characteristics and develop effective methods for noise reduction. The data is and converted to spectrograms using Short-Time Fourier Transform (STFT) for feature extraction. The resulting spectrograms are then used as input to the neural network. In general, the research aims to contribute to the development of improved speech recognition and processing technologies. Training and Validation Phases: The training phase is a critical step in developing a speech recognition system. It involves training a neural network model, specifically a combination of recurrent neural network (RNN) variants and neural network (CNN), to recognize speech patterns and convert them into text. The training process continues until the model achieves satisfactory accuracy on the training data. Following the training phase, the validation/testing phase is conducted to evaluate the system's performance on unseen data. A separate validation data set, derived from the training data, is used for this purpose. The goal of the validation phase is to monitor the system's performance and prevent overfitting, which refers to a situation where the model performs well on the training data but poorly on new, unseen data. The accuracy of the predicted transcription in the validation data is measured using the Word Error Rate (WER) metric. The model's hyperparameters, such as the learning rate, number of layers, and number of neurons, are adjusted during the validation phase to improve accuracy on the validation data. 3. System Description The proposed model adopts a hybrid deep learning architecture tailored for robust end-to-end Amharic Automatic Speech Recognition (ASR), particularly in noisy environments. It leverages the representational power of Convolutional Neural Networks (CNNs) for feature extraction, the sequential modeling capacity of Bidirectional Gated Recurrent Units (BiGRUs) for temporal dynamics, and the alignment flexibility of Connectionist Temporal Classification (CTC) loss for transcription without frame-level alignment. Input Representation: The model ingests time-frequency spectrograms derived from raw audio signals via Short-Time Fourier Transform (STFT). These spectrograms, preserving critical acoustic patterns, serve as the foundational input for subsequent neural processing. Feature Extraction via CNN: The spectrogram input is first passed through a cascade of 2D convolutional layers, each followed by batch normalization and ReLU activation. These layers are designed to capture local spectral and temporal patterns while enhancing the model’s invariance to distortions caused by noise or pronunciation variations. Temporal Modeling via BiGRU: The extracted feature maps are reshaped and fed into a stack of Bidirectional GRU layers. Each BiGRU layer consists of forward and backward GRU units that collectively capture long-range dependencies in both temporal directions. This bidirectional processing allows the model to contextualize each time frame with its preceding and succeeding acoustic context—crucial for accurate transcription in tonal and morphologically rich languages like Amharic. Dense Projection Layer: The final BiGRU output is passed through a fully connected (dense) layer with a linear activation, projecting the contextualized sequence into a high-dimensional space representing character probabilities. Sequence Alignment via CTC: To enable alignment-free training, the model employs Connectionist Temporal Classification (CTC) loss. This loss function facilitates the direct mapping from input sequences to target label sequences without requiring pre-segmented or frame-aligned annotations, making it highly effective for real-world ASR tasks where such alignment is infeasible. The complete model pipeline is visualized in Figure 2, illustrating the end-to-end flow from spectrogram input to character-level transcription. 3.1 Model Implementation This study describes This study presents an end-to-end automatic speech recognition (ASR) model for the Amharic language, built using TensorFlow/Keras. The architecture integrates convolutional layers, bidirectional recurrent layers, and dense output layers, trained with Connectionist Temporal Classification (CTC) loss to enable alignment-free transcription. Input Representation and Preprocessing: The input to the model consists of variable-length spectrograms generated from raw speech waveforms. These spectrograms, capturing temporal and frequency dynamics, are reshaped to 3D tensors with shape Convolutional Feature Extraction: The CNN block comprises two sequential 2D convolutional layers designed to extract hierarchical acoustic features. Each convolutional layer is followed by Batch Normalization and ReLU activation, facilitating faster convergence and improved generalization. After feature extraction, the output is reshaped into a 2D sequence, merging the spatial dimensions to prepare for recurrent processing. Temporal Modeling with Bidirectional GRU: The reshaped features are passed through a stack of Bidirectional Gated Recurrent Units (BiGRU). Each layer employs 512 units per direction and is configured with reset-after gating, tanh activations, and dropout (rate = 0.5) to mitigate overfitting. The bidirectional design allows the model to capture both forward and backward temporal dependencies in the speech signal, enhancing robustness in phoneme disambiguation and linguistic context modeling. Dense and Output Layers: The recurrent outputs are projected via a fully connected dense layer, expanding feature dimensionality for classification. A ReLU activation follows, succeeded by a dropout layer (rate = 0.5) to enforce regularization. The output layer applies a final dense transformation with a softmax activation, producing a probability distribution over the vocabulary (including an additional blank token required for CTC alignment). Sequence Decoding with CTC Loss: The model is trained using the CTC loss function, which enables learning without explicit frame-wise alignment between the audio input and text labels. This loss function handles variable-length sequences and supports flexible insertions of blank tokens for pause modeling and label separation. Decoding is performed using greedy or beam search strategies to generate the most probable character sequence. 3.2 Model Compilation The model is configured for training using TensorFlow/Keras’ compile method. First, the Adam optimizer is selected with a learning rate of 1×10 -4 for its proven effectiveness in optimizing deep neural networks. Next, the Connectionist Temporal Classification (CTC) loss is applied to accommodate the variable-length nature of both input speech signals and corresponding text transcriptions—a methodology widely adopted in ASR tasks (Lee & Watanabe, 2021). In practice, the compile method is invoked with these two core components, where the optimizer parameter is set to the Adam optimizer and the loss parameter is assigned the CTC loss function. The CTC mechanism inherently manages sequence alignment by incorporating a designated “blank” label between adjacent outputs, thereby eliminating the need for pre-segmented data. Let y be the ground truth label sequence, x be the input sequence, and y* be the predicted label sequence. The CTC loss function is defined as: During training, the model leverages a Connectionist Temporal Classification (CTC) layer (denoted as "ctc_loss") to calculate the probability of an output sequence given its corresponding input. The CTC loss function quantitatively measures the discrepancy between the predicted sequence and the ground truth labels, thereby guiding the network to learn a robust mapping from audio features to textual transcription. Minimization of this loss is the primary training objective. In this framework, the CTC layer ingests both the predicted probability tensor and the ground truth labels, computes the loss internally, and outputs this value, ensuring the model can handle variable-length sequences without the need for explicit frame-level alignment. Subsequent to training, the decoding stage transforms the predicted probability distributions into final textual outputs. This is accomplished using a beam search algorithm via TensorFlow’s K.ctc_decode function. Specifically, the decoding function initializes an array with the lengths of input sequences and iterates over each batch element, applying beam search (with greedy=False and a specified beam_width) to generate the most likely sequence. The decoded numerical labels are then converted into strings using a predefined mapping function (num_to_char), where TensorFlow’s tf.strings.reduce_join aggregates the individual characters into coherent text. The resulting transcription for each input is collected into a results list and returned as the final decoded output. This two-step process—first training with the CTC loss to manage alignment-free sequence prediction and then decoding via beam search—ensures that the model can robustly transcribe speech signals with variable lengths and ambiguous temporal structures. 3.3 Evaluation Metric The Word Error Rate (WER) is a standard metric for assessing the performance of speech recognition systems. It quantifies the accuracy of transcriptions by comparing the system's output to a reference transcription, taking into account three types of errors: substitutions, deletions, and insertions. WER is computed as the ratio of the total number of these errors to the total number of words in the reference text. Lower WER values indicate higher transcription accuracy and are instrumental in model comparison and hyperparameter tuning. WER= (I + D + S)/T where: I: insertion, D: deletion, S: substitution T: Total number of words 4. Results Several We conducted extensive experiments to assess the robustness of our model using 44 hours of noisy speech data. To counteract the detrimental effects of background noise, we investigated the impact of several noise reduction techniques: spectral subtraction, subspace filtering, and a combined approach. Spectral Subtraction: This method suppresses background noise by estimating and subtracting its spectral contribution from the noisy signal. Subspace Filtering: This technique further refines the signal by isolating its dominant subspace, thereby reducing noise influence. Combined Approach: Integrating both methods yielded the best performance. Our evaluation employed the Word Error Rate (WER) as the primary metric, along with signal-to-noise ratio (SNR) analysis to quantify the relative strength of the speech signal against the noise. Without applying any noise reduction, the system exhibited a WER of 10.5%. The introduction of spectral subtraction reduced the WER to 8.5%, while subspace filtering further improved accuracy with a WER of 7.4%. Notably, the combined noise reduction strategy achieved the lowest WER at 7.02% (see Table 1). Table 1 presents the Word Error Rate (WER) results Noise reduction techniques Noisy Data Spectral Subtraction WER = 8.5% Subspace Filtering WER =7.4% Combined (Spectral+Subspace) WER = 7.17% Without Noise- reduction technique WER = 10.5% The performance of the speech recognition system was quantified using the Word Error Rate (WER)—the percentage of words misrecognized relative to the reference transcription. Across a comprehensive dataset, our system achieved an overall WER of 7%. This evaluation encompasses multiple environmental scenarios, providing a robust analysis of the system's performance under diverse acoustic conditions. Our findings deliver critical insights into speech characteristics in various noise contexts, and they may guide the development of future technologies aimed at enhancing ASR accuracy in real-world settings. Furthermore, the analysis underscores the necessity for ongoing research into environmental influences on speech recognition. An illustration of the loss behavior on noisy data is provided in Figure 3. 5. Discussion In our end-to-end speech recognition framework, the training progress is typically visualized by plotting the loss against the number of training epochs—where each epoch denotes a full pass over the training dataset. In this context, the x-axis reflects the iterative training cycles, while the y-axis quantifies the loss, a metric indicating the divergence between the network's predictions and the true labels. At the onset of training, the model's loss is generally high due to limited exposure to the data. However, as training progresses, the network gradually internalizes the underlying patterns within the data, leading to a steady decline in the loss until it approaches a near-constant value. This convergence suggests that the model has effectively learned the mapping from inputs to outputs. It is important to note that while the training loss offers an indication of how well the model fits the training data, it does not necessarily reflect its performance on unseen data. To address this, a separate validation loss is monitored using a hold-out dataset that is not incorporated in the training process. The comparison between training and validation losses serves as an essential diagnostic tool: a continuously decreasing training loss accompanied by a rising validation loss may indicate overfitting, whereas persistently high losses on both datasets might suggest underfitting. By jointly analyzing these loss curves, researchers can obtain meaningful insights into the model’s learning dynamics and generalization capabilities. Adjustments to the network architecture or training hyperparameters are then made accordingly to optimize performance. A visual summary of these findings, which also includes the Word Error Rate (WER) for noisy data and predictions on unseen validation samples, is provided in Fig. 4 . Overall, our experimental results demonstrate that integrating noise reduction techniques—specifically spectral subtraction and subspace filtering—substantially enhances the performance of the Amharic ASR system when processing noisy inputs. For optimization, we employed the Adam optimizer, which is particularly well-suited for large-scale datasets and complex models (Graves & Jaitly, 2014). Adam effectively combines the strengths of AdaGrad and RMSProp by adapting the learning rate individually for each parameter. This mechanism promotes efficient convergence and stable training dynamics. In our experiments, we set the learning rate to 0.0001 to balance rapid learning with the precision required to avoid overshooting optimal parameter updates. Additionally, to mitigate overfitting and bolster the model's generalization capacity, a dropout rate of 0.5 was applied throughout the network (excluding the final output layer). This regularization strategy, which randomly deactivates half of the units during training, discourages the network from becoming overly dependent on specific features, thereby contributing to a more robust and accurate transcription performance in variable acoustic conditions. A comparison to select the best algorithm from those mentioned previously is compared in Table 2 below. Table 2 Comparison of BIGRU and BILSTM Model WER (%) Number of Parameters BIGRU 7.17 26,850,905 BILSTM 9.0 36,435,678 In this table, "Model" represents the type of architecture used (BIGRU and BILSTM), "WER (%)" denotes the word error rate achieved by each model, and "Number of Parameters" indicates the total number of parameters used in each model. We selected BIGRU because BILSTM required approximately twice the processing time of BIGRU. Furthermore, BILSTM requires a larger investment in computational resources. Our work has demonstrated the effectiveness of our end-to-end speech recognition models in a large amount of data, with the model achieving exceptional accuracy. These results can have important implications for a variety of applications, such as improving accessibility for individuals with hearing impairments or improving the accuracy of voice-controlled devices in controlled environments. 4.1 Error Analysis In our noisy dataset evaluation, the model produced a WER of 7%, with substitution and insertion errors being the predominant types. These errors are likely attributable to the inherent challenges posed by background noise and audio distortions, which impede the model’s ability to accurately decode the speech signal. A detailed examination of the mis-transcriptions revealed systematic character-level inaccuracies, notably instances where visually similar characters were confused. For example, the character "ከ" (ke) was frequently misrecognized as "ቀ" (q'e), suggesting that the acoustic or visual similarities between these characters may be contributing to the error patterns. These findings underscore the need for further refinement in feature extraction and noise-robust classification, particularly for language-specific nuances in the Amharic script. These characters have similar visual representations in their spectrogram representation. As a result, the model sometimes mistakenly replaced instances of "ከ" with "ቀ" in its transcriptions, which could introduce inaccuracies. Similarly, another swap involved the characters "ተ" (te) and "ጠ" (t'e). These characters share similar visual features in their visual representation of their audio. Consequently, the model occasionally misinterpreted "ተ" as "ጠ" and vice versa, leading to incorrect transcriptions. Another notable swap occurred between the characters "ች" (ch') and "ጭ" (tch'). These characters bear similarity in terms of their visual structure in spectrograms and became a challenge to CNN. As a result, the model occasionally confused "ች" with "ጭ," resulting in errors in the transcribed text. 6. Conclusions This study presents an innovative end-to-end ASR solution for the Amharic language by employing a unified RNN-based architecture. Traditional ASR systems often rely on multiple isolated components—such as language, acoustic, and pronunciation models—that necessitate significant manual effort and may introduce performance constraints. In contrast, our approach leverages the power of RNNs to simplify the speech recognition pipeline, reducing dependency on elaborate dictionaries and multi-stage processes. Our experiments, conducted on a comprehensive noisy dataset, reveal that incorporating noise reduction techniques—specifically spectral subtraction and subspace filtering—substantially enhances transcription accuracy. Notably, subspace filtering delivered marginally better noise suppression compared to spectral subtraction, with the combined approach resulting in a competitive WER of 7%. This robust performance under adverse acoustic conditions underscores the potential of our framework for practical, real-world applications. The adoption of BiGRU over BiLSTM was motivated by efficiency considerations, as BiGRU demonstrated shorter processing times and lower computational demands without compromising accuracy. These design choices, along with the integration of CNNs for automatic feature extraction and CTC for alignment-free sequence mapping, collectively contribute to an ASR system that is both efficient and highly adaptable. Beyond advancing speech recognition for a low-resource language like Amharic, our findings bear significant implications for the broader field. The streamlined architecture not only reduces development overhead and manual intervention but also paves the way for applications in voice search, voice commands for smart devices, customer interactions, and various other sectors. Future research should focus on further parameter optimization and testing the framework in diverse acoustic environments to enhance robustness and generalizability. In summary, our end-to-end RNN-based ASR model exhibits superior accuracy and resilience in noisy conditions, marking a substantial improvement over traditional approaches and providing a viable pathway for more accessible and scalable speech recognition systems. Declarations Author Contribution Y.A.E., T.T.A. and SAT, conceived and designed the study, wrote the main manuscript, and oversaw the experimental work. Y.A.E. led the model development, data analysis, and figure preparation, while T.T.A. and SAT contributed to the interpretation of the results and critical revisions of the manuscript. All authors reviewed and approved the final version of the manuscript. References Abate, S. T. (2005). Automatic Speech Recognition for Ahmaric. PhD dissertation. Hamburg University, Germany. Abebe Tsegaye (2019). Designing automatic speech recognition for Ge’ez language. Master’s thesis. Bahir Dar, Ethiopia. Azmeraw Dessalegn (2019). Syllable-based speaker-independent continuous speech recognition for Afan Oromo. Master’s thesis. Bahir Dar, Ethiopia. Bahl, L., Brown, P., De Souza, P. V., and Mercer, R. (1986). Maximum mutual information estimation of hidden Markov model parameters for speech recognition. In Acoustics, Speech, and Signal Processing, IEEE International Conference on ICASSP ’86 (pp. 49–52). doi: 10.1109/ICASSP.1986.1169179. Bisani, M., & Ney, H. (2005). Open vocabulary speech recognition with flat hybrid models. In INTERSPEECH (pp. 725–728). Bourlard, H. A., & Morgan, N. (1993). Connectionist Speech Recognition: A Hybrid Approach. Kluwer Academic Publishers, Norwell, MA, USA. ISBN 0792393961. Davis, S., & Mermelstein, P. (1980). Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. IEEE Transactions on Acoustics, Speech and Signal Processing, 28(4), 357–366. Baye, A., Tachbelie, Y., & Besacier, L. (2021). End-to-end Amharic speech recognition using LAS architecture. In Proceedings of the 12th Language Resources and Evaluation Conference (pp. 4058-4065). European Language Resources Association. Dumitru, C. O., & Gavat, I. (2006, June). A Comparative Study of Feature Extraction Methods Applied to Continuous Speech Recognition in Romanian Language. In Multimedia Signal Processing and Communications, 48th International Symposium ELMAR-2006 (pp. 115-118). IEEE. Furui, S., Ichiba, T., Shinozaki, T., Whittaker, E. W., & Iwano, K. (2005). Cluster-based modeling for ubiquitous speech recognition. Interspeech 2005, 2865-2868. Galescu, L. (2003). Recognition of out-of-vocabulary words with sub-lexical language models. In INTERSPEECH. Gebremedhin, Y. B., Duckhorn, F., Hoffmann, R., & Kraljevski, I. (2013). A new approach to develop a syllable-based, continuous Amharic speech recognizer. July, 1684–1689. Graves, A. (2012). Supervised Sequence Labelling with Recurrent Neural Networks, volume 385 of Studies in Computational Intelligence. Springer. Hebash H.O. Nasereddin, A. A. (January 2018). Classification techniques for automatic speech recognition (ASR) algorithms used with real-time speech translation. 2017 Computing Conference. London, UK: IEEE. Hinton, G., Deng, L., Yu, D., Dahl, G., Mohamed, A. r., Jaitly, N., Senior, A., Vanhoucke, V., Nguyen, P., Sainath, T., & Kingsbury, B. (2012). Deep neural networks for acoustic modeling in speech recognition. Signal Processing Magazine. Jaitly, N., & Hinton, G. E. (2011). Learning a better representation of speech soundwaves using restricted Boltzmann machines. In ICASSP (pp. 5884–5887). Pokhariya, J. S., & D. S. (2014). Sanskrit Speech Recognition using Hidden Markov Model Toolkit. International Journal of Engineering Research & Technology, 3(10). Kebebew, T. (2010). Speaker-dependent speech recognition for Afan Oromo using hybrid hidden Markov models and artificial neural network. Addis Ababa, Ethiopia: Addis Ababa University. Solomon Teferra Abate, W. M. (n.d.). Automatic Speech Recognition for an UnderResourced Language – Amharic. Department of Informatics, Natural Language Systems Group, University of Hamburg, Germany. Teferra, S., & Menzel, W. (2007). Syllable-based speech recognition for Amharic. June, 33–40. Tohye, T. G. (2015). Towards Improving the Performance of Spontaneous Amharic Speech Recognition. Master's thesis. Addis Ababa University, Addis Ababa, Ethiopia. Yifru, M. (2003). Application of Amharic Speech Recognition System to Command. Master's thesis. Addis Ababa University, Addis Ababa, Ethiopia. Graves, A., Fernández, S., Gomez, F., & Schmidhuber, J. (2006). Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks. In ICML, pages 369–376. ACM. Jurafsky, D., & Martin, J. H. (2009). Speech and Language Processing (2nd ed.). Prentice Hall. Huang, X., Acero, A., & Hon, H. W. (2001). Spoken Language Processing: A Guide to Theory, Algorithm, and System Development. Prentice Hall PTR. Graves, A., Mohamed, A.-r., & Hinton, G. (2013). Speech Recognition with Deep Recurrent Neural Networks. In IEEE International Conference on Acoustics, Speech and Signal Processing (pp. 6645–6649). Hinton, G., Deng, L., Yu, D., Dahl, G. E., Mohamed, A.-r., Jaitly, N., Senior, A., Vanhoucke, V., Nguyen, P., Sainath, T. N., & Kingsbury, B. (2012). Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups. IEEE Signal Processing Magazine, 29(6), 82–97. Graves, A., & Jaitly, N. (2014). Towards End-to-End Speech Recognition with Recurrent Neural Networks. In International Conference on Machine Learning (pp. 1764–1772). Jelinek, F., & Mercer, R. L. (1980). Interpolated Estimation of Markov Source Parameters from Sparse Data. In Proceedings of the Workshop on Pattern Recognition in Practice (pp. 381–397). Deng, L., & Yu, D. (2014). Deep Learning: Methods and Applications. Foundations and Trends in Signal Processing. Young, S., Evermann, G., Gales, M., Hain, T., Kershaw, D., Liu, X., Moore, G., Odell, J., Ollason, D., Povey, D., Valtchev, V., & Woodland, P. (2006). The HTK Book (for HTK Version 3.4). Cambridge University Engineering Department. Abdel-Hamid, O., & Jiang, H. (2013). Fast speaker adaptation of deep neural networks using model compression. In Proceedings of the 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (pp. 7822-7826). IEEE. Chavan, S., & Sable, P. (2013). Speech recognition using hidden Markov model. International Journal of Engineering Research and Applications, 3(4), 2319-2324. Novoa, J., Lleida, E., & Hernando, J. (2018). A comparative study of HMM-based and DNN-based ASR systems for under-resourced languages. Computer Speech & Language, 47, 1-22. Palaz, D., Kılıç, R., & Yılmaz, E. (2019). A comparative study of deep neural network and hidden Markov model based acoustic models for Turkish speech recognition. Journal of King Saud University-Computer and Information Sciences. Lee, K., & Kim, H. (2019). Dynamic Wrapping for Robust Speech Recognition in Adverse Environments. IEEE Access, 7, 129541-129550. Wang, Y., & Wang, D. (2019). A novel dynamic wrapping method for robust speech recognition under noisy environments. Journal of Ambient Intelligence and Humanized Computing, 10(5), 1825-1834. Chorowski, J. K., Bahdanau, D., Serdyuk, D., Cho, K., & Bengio, Y. (2015). Attention-based models for speech recognition. In Advances in Neural Information Processing Systems (pp. 577-585). Kim, J., Park, S., Kang, K., & Lee, H. (2017). Joint CTC-attention based end-to-end speech recognition using multi-task learning. Zhang, Y., Deng, K., Cao, S., & Ma, L. (2021). Improving Hybrid CTC/Attention End-to-end Speech Recognition with Pretrained Acoustic and Language Model. Deng, K., Zhang, Y., Cao, S., & Ma, L. (2022). Improving CTC-based speech recognition via knowledge transferring from pre-trained language models. Chen, C. (2021). ASR Inference with CTC Decoder. PyTorch. Hannun, A., Case, C., Casper, J., Catanzaro, B., Diamos, G., Elsen, E., … & Prenger, R. (2014). Deep speech: Scaling up end-to-end speech recognition. Igor Macedo Quintanilha (2017). End-to-end speech recognition applied to Brazilian Portuguese using deep learning. Hanan Aldarmaki, Asad Ullah, and Nazar Zaki (2021). Unsupervised Automatic Speech Recognition: A Review. arXiv.org, 2021, arxiv.org/abs/2106.04897. Abdelrahma Ahmed, Yasser Hifny, Khaled Shaalan, and Sergio Tolan (2016). Lexicon Free Arabic Speech Recognition. link.springer.com/chapter/10.1007/978-3-319-48308-5_15. Wren, Y., Titterington, J., & White, P. (2021). How many words make a sample? Determining the minimum number of word tokens needed in connected speech samples for child speech assessment. Clinical Linguistics & Phonetics, 35(8), 761-778. https://doi.org/10.1080/02699206.2020.1827458. Malik, M., Malik, M. K., Mehmood, K., & Makhdoom, I. (2021). Automatic speech recognition: a survey. Multimedia Tools and Applications, 80(6), 9411-9457. https://doi.org/10.1007/s11042-020-10073-7. Zhang, X., & Wang, H. (2011). A syllable-based connected speech recognition system for Mandarin. In 2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 4856-4859). IEEE. https://doi.org/10.1109/ICASSP.2011.5947265. Khan, A. I., & Zahid, S. (2012). A comparative study of speech recognition techniques for Urdu language. International Journal of Speech Technology, 15(4), 497-504. http://doi.org/10.1007/s10772-012-9175-z. Chen, C.-H., & Chen, C.-W. (2010). A novel speech recognition system based on a self-organizing feature map and a fuzzy expert system. Microprocessors and Microsystems, 34(6), 242-251. http://doi.org/10.1016/j.micpro.2010.04.003 Ge, X., Wu, L., Xia, D., & Zhang, P. (2013). A multi-objective HSA of nonlinear system modeling for hot skip-passing. In 2013 Sixth International Conference on Advanced Computational Intelligence (ICACI) (pp. 1-6). IEEE. https://doi.org/10.1109/ICACI.2013.6748484 Villa, A. E. P., Masulli, P., & Pons Rivero, A. J. (Eds.). (2016). Artificial neural networks and machine learning – ICANN 2016: 25th International Conference on Artificial Neural Networks, Barcelona, Spain, September 6-9, 2016, Proceedings, Part I. Springer International Publishing. https://doi.org/10.1007/978-3-319-44781-0 Chen, X., & Hu, Y. (2012). Learning optimal warping window size of DTW for time series classification. In 2012 11th International Conference on Information Science, Signal Processing, and their Applications (ISSPA) (pp. 1-4). IEEE. https://doi.org/10.1109/ISSPA.2012.6310488 La Rosa, M., Loos, P., & Pastor, O. (Eds.). (2016). Business Process Management: 14th International Conference, BPM 2016, Rio de Janeiro, Brazil, September 18-22, 2016. Proceedings. Springer. https://doi.org/10.1007/978-3-319-45348-4 Nguyen, V. N. (2016). A framework for business process improvement: A case study of a Norwegian manufacturing company (Master’s thesis). University of Stavanger. https://brage.bibsys.no/xmlui/bitstream/id/430871/16-00685-3%20Van%20Nhan%20Nguyen%20-%20Master’s%20Thesis.pdf%20266157_1_1.pdf LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. https://doi.org/10.1109/5.880079 Ientilucci, E. J. (2006). Statistical models for physically derived target sub-spaces. In S. S. Shen & P. E. Lewis (Eds.), Imaging Spectrometry XI (Vol. 6302, p. 63020A). SPIE. https://doi.org/10.1117/12.679525 Begel, A., & Bosch, J. (2013). The DevOps phenomenon. Communications of the ACM, 56(11), 44-49. http://dl.acm.org/citation.cfm?id=2113113.2113141 Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. Scherzer, O. (Ed.). (2015). Handbook of mathematical methods in imaging (2nd ed.). Springer. https://doi.org/10.1007/978-3-319-46478-7 Persagen Consulting. (n.d.). Machine learning. https://persagen.com/files/ml.html Saha, S. K., & Tripathy, A. K. (2017). Automated IT system failure prediction: A deep learning approach. International Journal of Computer Applications, 159(9), 1–6. https://www.researchgate.net/publication/313456329_Automated_IT_system_failure_prediction_A_deep_learning_approach Kotera, J., Šroubek, F., & Milanfar, P. (2013). Blind deconvolution using alternating maximum a posteriori estimation with heavy-tailed priors. In A. Petrosino, L. Maddalena, & P. Soda (Eds.), Computer analysis of images and patterns (pp. 59–66). Springer. https://doi.org/10.1007/978-3-642-40246-3_8 Artificial intelligence. (2022, January 5). In Wikipedia. https://en.wikipedia.org/wiki/Artificial_intelligence Mwiti, D. (2019, September 4). A 2019 guide for automatic speech recognition. Heartbeat. https://heartbeat.fritz.ai/a-2019-guide-for-automatic-speech-recognition-f1e1129a141c Persagen. (n.d.). Machine learning. Retrieved January 6, 2022, from https://persagen.com/files/ml.html Klein, A., & Kienle, A. (2012). Supporting distributed software development by modes of collaboration. Multimedia Systems, 18(6), 509–520. http://doi.org/10.1007/s00530-012-0266-0 Chen, Y., & Cong, J. (2017). DAPlace: Data-aware three-dimensional placement for large-scale heterogeneous FPGAs. In 2017 IEEE/ACM International Conference on Computer-Aided Design (ICCAD) (pp. 236–243). IEEE. http://doi.org/10.1109/ICCAD.2017.8203812 Zhang, Y., Zhang, H., Lin, H., & Zhang, Y. (2017). A new method for remote sensing image registration based on SURF and GMS. Remote Sensing, 9(3), 298. http://doi.org/10.3390/rs9030298 Schmidhuber, J. (2000). How to count time: A simple and general neural mechanism. ftp://ftp.idsia.ch/pub/juergen/TimeCount-IJCNN2000.pdf Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. https://arxiv.org/pdf/1406.1078v3.pdf Srivastava, A. (2018, November 14). Basic architecture of RNN and LSTM. PyDeepLearning. https://pydeeplearning.weebly.com/blog/basic-architecture-of-rnn-and-lstm Graves, A., Mohamed, A.-R., & Hinton, G. (2013). Speech recognition with deep recurrent neural networks. 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 6645–6649. https://www.cs.toronto.edu/~graves/asru_2013.pdf Zhang, Y., Zhang, J., Zhang, J., & Li, H. (2020). A survey on deep learning for named entity recognition. IEEE Transactions on Knowledge and Data Engineering, 32(9), 1635–1658. https://doi.org/10.1109/TKDE.2019.2906425 Shaik Johny Basha; Duggineni Veeraiah; Boddu Venkat Charan; Wiltrud Sahithi Yeddu; Devalla Ganesh Babu (2023) Olah, C. (2015, August 27). Understanding LSTM networks. http://colah.github.io/posts/2015-08-Understanding-LSTMs/ Ejigu, Y.A.; Asfaw, T.T. Enhancing Amharic Speech Recognition in Noisy Conditions through End-to-End Deep Learning. Preprints 2024, 2024020754. https://doi.org/10.20944/preprints202402.0754.v1 Ejigu, Y.A.; Asfaw, T.T. Large Scale Speech Recognition for Low Resource Language Amharic, an End-to-End Approach. Preprints 2024, 2024020813. https://doi.org/10.20944/preprints202402.0813.v1 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 24 Dec, 2025 Read the published version in International Journal of Speech Technology → Version 1 posted Editorial decision: Revision requested 18 Oct, 2025 Reviews received at journal 02 Sep, 2025 Reviewers agreed at journal 07 Aug, 2025 Reviewers invited by journal 29 Jul, 2025 Editor assigned by journal 19 Apr, 2025 Submission checks completed at journal 19 Apr, 2025 First submitted to journal 10 Apr, 2025 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-6419413","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":494026801,"identity":"020cd453-b8aa-4f5a-a68e-22c79016efc5","order_by":0,"name":"Yohannes Ayana Ejigu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBUlEQVRIiWNgGAWjYFACHmSOgYQciDrwgHgtBTbGYC0JxGv5kJbYAKLxaTE43nvwc+EOG3v+GcnPHvwwOJw+P+zwQ6AtdnK6DTi0nDmXLD3zTFrijBtp5oY9BodzN95OMwBqSTY2O4BDy40cA2netsMJDDcSzCR4QFpmJ4C0HEjchluL8W/etv/28jfSv0n+ATrMcHb6B0JazIC2HGDcAGLwGKQlyEvn4LdF8swZM2vetuTEjWfelEnLGNgYbpDOKTiQYIDbL3zHe4xv87bZ2csdT98m+eaPhLz87PTNHz5U2Mnh0qIAFxdIgDoVLGKAXTkIyDfAWPwH0EVGwSgYBaNgFEAAAMFZZkcyr+mZAAAAAElFTkSuQmCC","orcid":"","institution":"Bahir Dar University","correspondingAuthor":true,"prefix":"","firstName":"Yohannes","middleName":"Ayana","lastName":"Ejigu","suffix":""},{"id":494026802,"identity":"14d6a1fe-0da8-40a7-9055-89a82dafd21c","order_by":1,"name":"Tesfa Tegegne Asfaw","email":"","orcid":"","institution":"Bahir Dar University","correspondingAuthor":false,"prefix":"","firstName":"Tesfa","middleName":"Tegegne","lastName":"Asfaw","suffix":""},{"id":494026804,"identity":"1caa097a-d635-492b-92c6-306f8931b16c","order_by":2,"name":"Surafel Amsalu Tadesse","email":"","orcid":"","institution":"Bahir Dar University","correspondingAuthor":false,"prefix":"","firstName":"Surafel","middleName":"Amsalu","lastName":"Tadesse","suffix":""}],"badges":[],"createdAt":"2025-04-10 10:53:21","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6419413/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6419413/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s10772-025-10224-x","type":"published","date":"2025-12-24T15:57:59+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":88239722,"identity":"1993c8ca-b882-40cb-8c71-dc003eec3bbe","added_by":"auto","created_at":"2025-08-04 11:03:10","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":253916,"visible":true,"origin":"","legend":"\u003cp\u003epresents the top 3 rows of the noisydata\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6419413/v1/5621b1ee123151814a62ba4a.png"},{"id":88240647,"identity":"261bcf0f-be9a-4e1a-9552-9887ff9dee1c","added_by":"auto","created_at":"2025-08-04 11:11:10","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":526753,"visible":true,"origin":"","legend":"\u003cp\u003eProposed architecture of the model\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6419413/v1/07e242b34acb10fd8e4f42ca.png"},{"id":88239728,"identity":"d5917916-7d63-458a-8369-b8c9de5893fe","added_by":"auto","created_at":"2025-08-04 11:03:10","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":172770,"visible":true,"origin":"","legend":"\u003cp\u003epresents the loss function in a noisy data\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6419413/v1/563e211a975c7a84aef8fbb0.png"},{"id":88241140,"identity":"8eaab587-5ca3-455b-b314-3fd2d13856e4","added_by":"auto","created_at":"2025-08-04 11:19:11","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":746038,"visible":true,"origin":"","legend":"\u003cp\u003eWER for noisy data and the predictions for unseen validation data via RTX800 NVIDIA GPU.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6419413/v1/cf7da42668204a73d0c05656.png"},{"id":99172491,"identity":"47f2a5f0-3b23-48c1-ad54-e0e7170402a7","added_by":"auto","created_at":"2025-12-29 16:10:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2115617,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6419413/v1/b6388d17-42ae-40aa-983e-69071eca850d.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Noise-Robust End-to-End Framework for Amharic Speech Recognition","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eSpeech recognition\u0026mdash;often referred to as automatic speech recognition (ASR), computer speech recognition, or speech-to-text\u0026mdash;facilitates the conversion of spoken language into written text. While it is sometimes conflated with voice recognition, which identifies the speaker rather than the words themselves, ASR focuses exclusively on the transcription of spoken utterances. Recent advancements in deep learning have prompted a surge of applications across multiple sectors, from healthcare and finance to public safety, providing immense time-saving and life-saving potential (Hanan Aldarmaki, Asad Ullah, \u0026amp; Nazar Zaki, 2021).\u003c/p\u003e\u003cp\u003eOver the years, a variety of methods have been explored for constructing ASR systems, including Dynamic Time Warping (DTW), Hidden Markov Models (HMMs), Dynamic Bayesian Networks (DBNs), and artificial neural networks (ANNs) (Hebash H.O. Nasereddin, January 2018). In particular, HMMs have historically demonstrated effective performance in decoding acoustic signals into phonetic states (Kebebew, 2010). However, these approaches often rely on multiple, separately trained modules\u0026mdash;such as pronunciation dictionaries, acoustic models, and language models\u0026mdash;that must work in tandem. This not only increases computational and engineering complexity but can also introduce inconsistencies between frame-level training objectives and sequence-level performance (Bourlard \u0026amp; Morgan, 1993; Hinton et al., 2012). For instance, improvements in per-frame accuracy sometimes fail to boost overall transcription accuracy, highlighting a fundamental mismatch in training and inference objectives (Bahl et al., 1986; Jaitly et al., 2012). Furthermore, traditional pipeline systems typically require meticulous human curation of phonemic dictionaries to map words into subword units, introducing significant time and resource overhead (Graves, A., \u0026amp; Jaitly, N., 2014). For languages such as Amharic\u0026mdash;where large-scale lexical databases may not be readily available or are resource-intensive to build\u0026mdash;these dependencies present substantial barriers. Conventional speech recognition systems for Amharic often ignore noisy real-world conditions and focus on relatively clean datasets, thereby limiting their applicability (Baye, A., Tachbelie, Y., \u0026amp; Besacier, L., 2021).\u003c/p\u003e\u003cp\u003eTo address these challenges, we propose an end-to-end ASR framework that incorporates a recurrent neural network (RNN) as the principal component, streamlining the entire speech-recognition pipeline. By replacing separate acoustic and pronunciation modules with a single neural sequence model, our approach bypasses the need for extensive dictionary construction and aligns model training directly with the sequence-level transcription goal. Although direct waveform-to-text transcription has been explored (Graves, 2012; Jaitly \u0026amp; Hinton, 2011), computational overhead remains a concern, making spectrogram-based preprocessing a practical compromise. Additionally, to handle dynamic noise environments, we incorporate robust feature extraction and noise-mitigation strategies into our end-to-end pipeline.\u003c/p\u003e\u003cp\u003eGiven that current end-to-end architectures for Amharic often overlook noisy data and rarely employ deep neural methods for feature extraction (Baye et al., 2021), this research fills a crucial gap by providing a noise-aware, end-to-end ASR model for Amharic. In contrast to conventional HMM-based solutions, our approach eliminates multiple sub-pipelines\u0026mdash;such as manual pronunciation dictionaries\u0026mdash;thus reducing the reliance on domain experts. We also leverage Connectionist Temporal Classification (CTC) to address alignment issues that typically plague conventional systems.\u003c/p\u003e\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e\u003ch2\u003e1.1 Research Questions\u003c/h2\u003e\u003cp\u003eThis study addresses the following key questions:\u003c/p\u003e\u003cp\u003eRQ1: What impact does applying erosion to spectrogram features have on the accuracy of an Amharic speech recognition model?\u003c/p\u003e\u003cp\u003eRQ2: Which noise reduction approach\u0026mdash;spectral subtraction, subspace filtering, or a combined strategy\u0026mdash;proves most effective for end-to-end Amharic ASR?\u003c/p\u003e\u003cp\u003eRQ3: Between BiLSTM and BiGRU architectures, which yields superior performance and efficiency for Amharic ASR under noisy conditions?\u003c/p\u003e\u003c/div\u003e"},{"header":"2. Data","content":"\u003cp\u003eWe conducted audio recordings of individuals delivering speeches in the Sidama region, where Amharic is not their primary language. The noisy dataset consists of 400 distinct sentences, each sentence being read by 50 different individuals. In total, 20,000 sentences are used for the recordings. The noisy audios are read speech data which last from a minimum of 4 seconds to a maximum of 20 seconds. The total noisy data is around 44 hours and 46 minutes. The dataset referenced is currently accessible online under our authorship. It is readily available via the provided link below:\u003c/p\u003e\u003cp\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://drive.google.com/drive/folders/155kEo_yyJRIOyBTv8ruhHmHdNKQPx_My\u003c/span\u003e\u003cspan address=\"https://drive.google.com/drive/folders/155kEo_yyJRIOyBTv8ruhHmHdNKQPx_My\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eFor the speech recording (noisy data), a corresponding transcription was created. This process involved carefully listening to the audio and converting it into written text. The text data serves as the ground truth for the deep learning model, allowing it to learn the relationship between the audio features and the corresponding text.\u003c/p\u003e\u003cp\u003eCapturing speech data in environments characterized by high levels of noise demands meticulous attention to hardware, software, and environmental variables to uphold the integrity and caliber of the dataset. In pursuit of accurate speech capture within such challenging conditions, we deliberately selected robust microphones engineered to excel in noise-laden settings, alongside the latest smartphone technology to enhance our recording capabilities. Our recording protocol was underpinned by the utilization of industry-standard software, affording us the capacity to faithfully capture the vocalizations of 50 distinct individuals. These software tools facilitated precise and comprehensive audio recording, ensuring the fidelity and authenticity of the dataset. Recording sessions were deliberately orchestrated across a spectrum of noisy environments, mirroring real-world scenarios with precision. These environments encompassed bustling urban locales, resonating with the clamor of vehicular traffic and bustling crowds, as well as tranquil public spaces such as parks, where the ambient soundscape embraced the melodic trills of avian inhabitants and the rustling of foliage. Additionally, recordings were made in various other noisy settings, including environments characterized by human activity at a distance.\u003c/p\u003e\u003cp\u003eThe decision to collect the corpus in the Sidama region is rooted in the area's significant linguistic and demographic characteristics. Recent studies (Shaik Johny Basha et al., 2023) report over 60,000 Amharic speakers in Ethiopia. Moreover, the Sidama region attracts individuals from diverse geographic backgrounds, fostering a rich linguistic landscape characterized by various dialects and linguistic influences. While we acknowledge the importance of capturing speech data from regions where Amharic is the primary language, we also recognize the necessity of ensuring geographic diversity within our dataset. Incorporating speakers from areas where Amharic serves as a second language is paramount for enhancing the robustness and diversity of our model. By including individuals from diverse linguistic backgrounds, we aim to improve the model's adaptability to different speech patterns and linguistic contexts, thereby strengthening its performance and usability in real-world applications.\u003c/p\u003e\u003cp\u003eThe text corpus comprises 44,937 unique words within the noisy Amharic language dataset. Additionally, the corpus encompasses more than 250 characters. To ensure linguistic balance among phonemes, a comprehensive frequency analysis of both words and characters has been conducted using a Python script. This analysis not only quantifies the occurrence of individual phonemes but also provides insights into the distribution and representation of linguistic elements within the corpus.\u003c/p\u003e\u003cp\u003eThe noisy environment in which the data was collected was deliberately designed to incorporate diverse sources of background noise. These included ambient sounds such as bird calls from a nearby park, traffic noise from passing cars, wind disturbances, and chatter from other individuals, such as conversations in a cafeteria setting.\u003c/p\u003e\u003cp\u003eFor this study, the demographic composition of the speakers primarily comprises young adults and middle-aged adults, with an equal representation of both male and female participants. It's important to note that our current research focus has led us to include a specific age range, spanning from 18 to 50 years, ensuring a balanced distribution across genders. The dataset was divided into three subsets for training, validation, and testing, with a split ratio of 80%, 10%, and 10%, respectively.\u003c/p\u003e\u003cp\u003eThe transcribed text paired with the file name of the audio is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e below.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe collected data is analyzed to gain insight into speech characteristics and develop effective methods for noise reduction. The data is and converted to spectrograms using Short-Time Fourier Transform (STFT) for feature extraction. The resulting spectrograms are then used as input to the neural network. In general, the research aims to contribute to the development of improved speech recognition and processing technologies.\u003c/p\u003e\u003cp\u003eTraining and Validation Phases: The training phase is a critical step in developing a speech recognition system. It involves training a neural network model, specifically a combination of recurrent neural network (RNN) variants and neural network (CNN), to recognize speech patterns and convert them into text. The training process continues until the model achieves satisfactory accuracy on the training data.\u003c/p\u003e\u003cp\u003eFollowing the training phase, the validation/testing phase is conducted to evaluate the system's performance on unseen data. A separate validation data set, derived from the training data, is used for this purpose. The goal of the validation phase is to monitor the system's performance and prevent overfitting, which refers to a situation where the model performs well on the training data but poorly on new, unseen data. The accuracy of the predicted transcription in the validation data is measured using the Word Error Rate (WER) metric. The model's hyperparameters, such as the learning rate, number of layers, and number of neurons, are adjusted during the validation phase to improve accuracy on the validation data.\u003c/p\u003e"},{"header":"3. System Description","content":"\u003cp\u003eThe proposed model adopts a hybrid deep learning architecture tailored for robust end-to-end Amharic Automatic Speech Recognition (ASR), particularly in noisy environments. It leverages the representational power of Convolutional Neural Networks (CNNs) for feature extraction, the sequential modeling capacity of Bidirectional Gated Recurrent Units (BiGRUs) for temporal dynamics, and the alignment flexibility of Connectionist Temporal Classification (CTC) loss for transcription without frame-level alignment.\u003c/p\u003e\n\u003cp\u003eInput Representation: The model ingests time-frequency spectrograms derived from raw audio signals via Short-Time Fourier Transform (STFT). These spectrograms, preserving critical acoustic patterns, serve as the foundational input for subsequent neural processing.\u003c/p\u003e\n\u003cp\u003eFeature Extraction via CNN: The spectrogram input is first passed through a cascade of 2D convolutional layers, each followed by batch normalization and ReLU activation. These layers are designed to capture local spectral and temporal patterns while enhancing the model\u0026rsquo;s invariance to distortions caused by noise or pronunciation variations.\u003c/p\u003e\n\u003cp\u003eTemporal Modeling via BiGRU: The extracted feature maps are reshaped and fed into a stack of Bidirectional GRU layers. Each BiGRU layer consists of forward and backward GRU units that collectively capture long-range dependencies in both temporal directions. This bidirectional processing allows the model to contextualize each time frame with its preceding and succeeding acoustic context\u0026mdash;crucial for accurate transcription in tonal and morphologically rich languages like Amharic.\u003c/p\u003e\n\u003cp\u003eDense Projection Layer: The final BiGRU output is passed through a fully connected (dense) layer with a linear activation, projecting the contextualized sequence into a high-dimensional space representing character probabilities.\u003c/p\u003e\n\u003cp\u003eSequence Alignment via CTC: To enable alignment-free training, the model employs Connectionist Temporal Classification (CTC) loss. This loss function facilitates the direct mapping from input sequences to target label sequences without requiring pre-segmented or frame-aligned annotations, making it highly effective for real-world ASR tasks where such alignment is infeasible.\u003c/p\u003e\n\u003cp\u003eThe complete model pipeline is visualized in Figure 2, illustrating the end-to-end flow from spectrogram input to character-level transcription.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.1 Model Implementation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study describes This study presents an end-to-end automatic speech recognition (ASR) model for the Amharic language, built using TensorFlow/Keras. The architecture integrates convolutional layers, bidirectional recurrent layers, and dense output layers, trained with Connectionist Temporal Classification (CTC) loss to enable alignment-free transcription.\u003c/p\u003e\n\u003cp\u003eInput Representation and Preprocessing: The input to the model consists of variable-length spectrograms generated from raw speech waveforms. These spectrograms, capturing temporal and frequency dynamics, are reshaped to 3D tensors with shape\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConvolutional Feature Extraction: The CNN block comprises two sequential 2D convolutional layers designed to extract hierarchical acoustic features.\u003c/p\u003e\n\u003cp\u003eEach convolutional layer is followed by Batch Normalization and ReLU activation, facilitating faster convergence and improved generalization. After feature extraction, the output is reshaped into a 2D sequence, merging the spatial dimensions to prepare for recurrent processing.\u003c/p\u003e\n\u003cp\u003eTemporal Modeling with Bidirectional GRU: The reshaped features are passed through a stack of Bidirectional Gated Recurrent Units (BiGRU). Each layer employs 512 units per direction and is configured with reset-after gating, tanh activations, and dropout (rate = 0.5) to mitigate overfitting. The bidirectional design allows the model to capture both forward and backward temporal dependencies in the speech signal, enhancing robustness in phoneme disambiguation and linguistic context modeling.\u003c/p\u003e\n\u003cp\u003eDense and Output Layers: The recurrent outputs are projected via a fully connected dense layer, expanding feature dimensionality for classification. A ReLU activation follows, succeeded by a dropout layer (rate = 0.5) to enforce regularization. The output layer applies a final dense transformation with a softmax activation, producing a probability distribution over the vocabulary (including an additional blank token required for CTC alignment).\u003c/p\u003e\n\u003cp\u003eSequence Decoding with CTC Loss: The model is trained using the CTC loss function, which enables learning without explicit frame-wise alignment between the audio input and text labels. This loss function handles variable-length sequences and supports flexible insertions of blank tokens for pause modeling and label separation. Decoding is performed using greedy or beam search strategies to generate the most probable character sequence.\u003c/p\u003e\n\u003ch3\u003e3.2 Model Compilation\u003c/h3\u003e\n\u003cp\u003eThe model is configured for training using TensorFlow/Keras\u0026rsquo; compile method. First, the Adam optimizer is selected with a learning rate of 1\u0026times;10\u003csup\u003e-4\u003c/sup\u003e for its proven effectiveness in optimizing deep neural networks. Next, the Connectionist Temporal Classification (CTC) loss is applied to accommodate the variable-length nature of both input speech signals and corresponding text transcriptions\u0026mdash;a methodology widely adopted in ASR tasks (Lee \u0026amp; Watanabe, 2021).\u003c/p\u003e\n\u003cp\u003eIn practice, the compile method is invoked with these two core components, where the optimizer parameter is set to the Adam optimizer and the loss parameter is assigned the CTC loss function. The CTC mechanism inherently manages sequence alignment by incorporating a designated \u0026ldquo;blank\u0026rdquo; label between adjacent outputs, thereby eliminating the need for pre-segmented data.\u003c/p\u003e\n\u003cp\u003eLet y be the ground truth label sequence, x be the input sequence, and y* be the predicted label sequence. The CTC loss function is defined as:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"884\" height=\"447\"\u003e\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003eDuring training, the model leverages a Connectionist Temporal Classification (CTC) layer (denoted as \u0026quot;ctc_loss\u0026quot;) to calculate the probability of an output sequence given its corresponding input. The CTC loss function quantitatively measures the discrepancy between the predicted sequence and the ground truth labels, thereby guiding the network to learn a robust mapping from audio features to textual transcription. Minimization of this loss is the primary training objective.\u003c/p\u003e\n\u003cp\u003eIn this framework, the CTC layer ingests both the predicted probability tensor and the ground truth labels, computes the loss internally, and outputs this value, ensuring the model can handle variable-length sequences without the need for explicit frame-level alignment.\u003c/p\u003e\n\u003cp\u003eSubsequent to training, the decoding stage transforms the predicted probability distributions into final textual outputs. This is accomplished using a beam search algorithm via TensorFlow\u0026rsquo;s K.ctc_decode function. Specifically, the decoding function initializes an array with the lengths of input sequences and iterates over each batch element, applying beam search (with greedy=False and a specified beam_width) to generate the most likely sequence. The decoded numerical labels are then converted into strings using a predefined mapping function (num_to_char), where TensorFlow\u0026rsquo;s tf.strings.reduce_join aggregates the individual characters into coherent text.\u003c/p\u003e\n\u003cp\u003eThe resulting transcription for each input is collected into a results list and returned as the final decoded output. This two-step process\u0026mdash;first training with the CTC loss to manage alignment-free sequence prediction and then decoding via beam search\u0026mdash;ensures that the model can robustly transcribe speech signals with variable lengths and ambiguous temporal structures.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3 Evaluation Metric\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Word Error Rate (WER) is a standard metric for assessing the performance of speech recognition systems. It quantifies the accuracy of transcriptions by comparing the system\u0026apos;s output to a reference transcription, taking into account three types of errors: substitutions, deletions, and insertions. WER is computed as the ratio of the total number of these errors to the total number of words in the reference text. Lower WER values indicate higher transcription accuracy and are instrumental in model comparison and hyperparameter tuning.\u003c/p\u003e\n\u003cp\u003eWER= (I + D + S)/T \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;where: I: insertion, D: deletion, S: substitution\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eT: Total number of words\u003c/p\u003e"},{"header":"4. Results","content":"\u003cp\u003eSeveral\u0026nbsp;We conducted extensive experiments to assess the robustness of our model using 44 hours of noisy speech data. To counteract the detrimental effects of background noise, we investigated the impact of several noise reduction techniques: spectral subtraction, subspace filtering, and a combined approach.\u003c/p\u003e\n\u003cp\u003eSpectral Subtraction: This method suppresses background noise by estimating and subtracting its spectral contribution from the noisy signal.\u003c/p\u003e\n\u003cp\u003eSubspace Filtering: This technique further refines the signal by isolating its dominant subspace, thereby reducing noise influence.\u003c/p\u003e\n\u003cp\u003eCombined Approach: Integrating both methods yielded the best performance.\u003c/p\u003e\n\u003cp\u003eOur evaluation employed the Word Error Rate (WER) as the primary metric, along with signal-to-noise ratio (SNR) analysis to quantify the relative strength of the speech signal against the noise. Without applying any noise reduction, the system exhibited a WER of 10.5%. The introduction of spectral subtraction reduced the WER to 8.5%, while subspace filtering further improved accuracy with a WER of 7.4%. Notably, the combined noise reduction strategy achieved the lowest WER at 7.02% (see Table 1).\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;1\u0026nbsp;presents\u0026nbsp;the\u0026nbsp;Word\u0026nbsp;Error\u0026nbsp;Rate\u0026nbsp;(WER) results\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 43.299%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eNoise reduction techniques\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 56.701%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eNoisy Data\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 43.299%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eSpectral Subtraction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 56.701%;\"\u003e\n \u003cp\u003eWER\u0026nbsp;= 8.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 43.299%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eSubspace Filtering\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 56.701%;\"\u003e\n \u003cp\u003eWER\u0026nbsp;=7.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 43.299%;\"\u003e\n \u003cp\u003eCombined (Spectral+Subspace)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 56.701%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eWER\u0026nbsp;= 7.17%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 43.299%;\"\u003e\n \u003cp\u003eWithout\u0026nbsp;Noise- reduction technique\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 56.701%;\"\u003e\n \u003cp\u003eWER\u0026nbsp;= 10.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe performance of the speech recognition system was quantified using the Word Error Rate (WER)\u0026mdash;the percentage of words misrecognized relative to the reference transcription. Across a comprehensive dataset, our system achieved an overall WER of 7%. This evaluation encompasses multiple environmental scenarios, providing a robust analysis of the system\u0026apos;s performance under diverse acoustic conditions.\u003c/p\u003e\n\u003cp\u003eOur findings deliver critical insights into speech characteristics in various noise contexts, and they may guide the development of future technologies aimed at enhancing ASR accuracy in real-world settings. Furthermore, the analysis underscores the necessity for ongoing research into environmental influences on speech recognition. An illustration of the loss behavior on noisy data is provided in Figure 3.\u003c/p\u003e"},{"header":"5. Discussion","content":"\u003cp\u003eIn our end-to-end speech recognition framework, the training progress is typically visualized by plotting the loss against the number of training epochs\u0026mdash;where each epoch denotes a full pass over the training dataset. In this context, the x-axis reflects the iterative training cycles, while the y-axis quantifies the loss, a metric indicating the divergence between the network's predictions and the true labels.\u003c/p\u003e\u003cp\u003eAt the onset of training, the model's loss is generally high due to limited exposure to the data. However, as training progresses, the network gradually internalizes the underlying patterns within the data, leading to a steady decline in the loss until it approaches a near-constant value. This convergence suggests that the model has effectively learned the mapping from inputs to outputs.\u003c/p\u003e\u003cp\u003eIt is important to note that while the training loss offers an indication of how well the model fits the training data, it does not necessarily reflect its performance on unseen data. To address this, a separate validation loss is monitored using a hold-out dataset that is not incorporated in the training process. The comparison between training and validation losses serves as an essential diagnostic tool: a continuously decreasing training loss accompanied by a rising validation loss may indicate overfitting, whereas persistently high losses on both datasets might suggest underfitting.\u003c/p\u003e\u003cp\u003eBy jointly analyzing these loss curves, researchers can obtain meaningful insights into the model\u0026rsquo;s learning dynamics and generalization capabilities. Adjustments to the network architecture or training hyperparameters are then made accordingly to optimize performance. A visual summary of these findings, which also includes the Word Error Rate (WER) for noisy data and predictions on unseen validation samples, is provided in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eOverall, our experimental results demonstrate that integrating noise reduction techniques\u0026mdash;specifically spectral subtraction and subspace filtering\u0026mdash;substantially enhances the performance of the Amharic ASR system when processing noisy inputs.\u003c/p\u003e\u003cp\u003eFor optimization, we employed the Adam optimizer, which is particularly well-suited for large-scale datasets and complex models (Graves \u0026amp; Jaitly, 2014). Adam effectively combines the strengths of AdaGrad and RMSProp by adapting the learning rate individually for each parameter. This mechanism promotes efficient convergence and stable training dynamics. In our experiments, we set the learning rate to 0.0001 to balance rapid learning with the precision required to avoid overshooting optimal parameter updates.\u003c/p\u003e\u003cp\u003eAdditionally, to mitigate overfitting and bolster the model's generalization capacity, a dropout rate of 0.5 was applied throughout the network (excluding the final output layer). This regularization strategy, which randomly deactivates half of the units during training, discourages the network from becoming overly dependent on specific features, thereby contributing to a more robust and accurate transcription performance in variable acoustic conditions.\u003c/p\u003e\u003cp\u003eA comparison to select the best algorithm from those mentioned previously is compared in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e below.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\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\u003eComparison of BIGRU and BILSTM\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eWER (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNumber of Parameters\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBIGRU\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e7.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e26,850,905\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBILSTM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e9.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e36,435,678\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn this table, \"Model\" represents the type of architecture used (BIGRU and BILSTM), \"WER (%)\" denotes the word error rate achieved by each model, and \"Number of Parameters\" indicates the total number of parameters used in each model. We selected BIGRU because BILSTM required approximately twice the processing time of BIGRU. Furthermore, BILSTM requires a larger investment in computational resources.\u003c/p\u003e\u003cp\u003eOur work has demonstrated the effectiveness of our end-to-end speech recognition models in a large amount of data, with the model achieving exceptional accuracy. These results can have important implications for a variety of applications, such as improving accessibility for individuals with hearing impairments or improving the accuracy of voice-controlled devices in controlled environments.\u003c/p\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Error Analysis\u003c/h2\u003e\u003cp\u003eIn our noisy dataset evaluation, the model produced a WER of 7%, with substitution and insertion errors being the predominant types. These errors are likely attributable to the inherent challenges posed by background noise and audio distortions, which impede the model\u0026rsquo;s ability to accurately decode the speech signal.\u003c/p\u003e\u003cp\u003eA detailed examination of the mis-transcriptions revealed systematic character-level inaccuracies, notably instances where visually similar characters were confused. For example, the character \"ከ\" (ke) was frequently misrecognized as \"ቀ\" (q'e), suggesting that the acoustic or visual similarities between these characters may be contributing to the error patterns. These findings underscore the need for further refinement in feature extraction and noise-robust classification, particularly for language-specific nuances in the Amharic script. These characters have similar visual representations in their spectrogram representation. As a result, the model sometimes mistakenly replaced instances of \"ከ\" with \"ቀ\" in its transcriptions, which could introduce inaccuracies. Similarly, another swap involved the characters \"ተ\" (te) and \"ጠ\" (t'e). These characters share similar visual features in their visual representation of their audio. Consequently, the model occasionally misinterpreted \"ተ\" as \"ጠ\" and vice versa, leading to incorrect transcriptions. Another notable swap occurred between the characters \"ች\" (ch') and \"ጭ\" (tch'). These characters bear similarity in terms of their visual structure in spectrograms and became a challenge to CNN. As a result, the model occasionally confused \"ች\" with \"ጭ,\" resulting in errors in the transcribed text.\u003c/p\u003e\u003c/div\u003e"},{"header":"6. Conclusions","content":"\u003cp\u003eThis study presents an innovative end-to-end ASR solution for the Amharic language by employing a unified RNN-based architecture. Traditional ASR systems often rely on multiple isolated components\u0026mdash;such as language, acoustic, and pronunciation models\u0026mdash;that necessitate significant manual effort and may introduce performance constraints. In contrast, our approach leverages the power of RNNs to simplify the speech recognition pipeline, reducing dependency on elaborate dictionaries and multi-stage processes.\u003c/p\u003e\u003cp\u003eOur experiments, conducted on a comprehensive noisy dataset, reveal that incorporating noise reduction techniques\u0026mdash;specifically spectral subtraction and subspace filtering\u0026mdash;substantially enhances transcription accuracy. Notably, subspace filtering delivered marginally better noise suppression compared to spectral subtraction, with the combined approach resulting in a competitive WER of 7%. This robust performance under adverse acoustic conditions underscores the potential of our framework for practical, real-world applications.\u003c/p\u003e\u003cp\u003eThe adoption of BiGRU over BiLSTM was motivated by efficiency considerations, as BiGRU demonstrated shorter processing times and lower computational demands without compromising accuracy. These design choices, along with the integration of CNNs for automatic feature extraction and CTC for alignment-free sequence mapping, collectively contribute to an ASR system that is both efficient and highly adaptable.\u003c/p\u003e\u003cp\u003eBeyond advancing speech recognition for a low-resource language like Amharic, our findings bear significant implications for the broader field. The streamlined architecture not only reduces development overhead and manual intervention but also paves the way for applications in voice search, voice commands for smart devices, customer interactions, and various other sectors. Future research should focus on further parameter optimization and testing the framework in diverse acoustic environments to enhance robustness and generalizability.\u003c/p\u003e\u003cp\u003eIn summary, our end-to-end RNN-based ASR model exhibits superior accuracy and resilience in noisy conditions, marking a substantial improvement over traditional approaches and providing a viable pathway for more accessible and scalable speech recognition systems.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eY.A.E., T.T.A. and SAT, conceived and designed the study, wrote the main manuscript, and oversaw the experimental work. Y.A.E. led the model development, data analysis, and figure preparation, while T.T.A. and SAT contributed to the interpretation of the results and critical revisions of the manuscript. All authors reviewed and approved the final version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAbate, S. T. (2005). Automatic Speech Recognition for Ahmaric. PhD dissertation. Hamburg University, Germany.\u003c/li\u003e\n \u003cli\u003eAbebe Tsegaye (2019). Designing automatic speech recognition for Ge\u0026rsquo;ez language. Master\u0026rsquo;s thesis. Bahir Dar, Ethiopia.\u003c/li\u003e\n \u003cli\u003eAzmeraw Dessalegn (2019). Syllable-based speaker-independent continuous speech recognition for Afan Oromo. Master\u0026rsquo;s thesis. Bahir Dar, Ethiopia.\u003c/li\u003e\n \u003cli\u003eBahl, L., Brown, P., De Souza, P. V., and Mercer, R. (1986). Maximum mutual information estimation of hidden Markov model parameters for speech recognition. In Acoustics, Speech, and Signal Processing, IEEE International Conference on ICASSP \u0026rsquo;86 (pp. 49\u0026ndash;52). doi: 10.1109/ICASSP.1986.1169179.\u003c/li\u003e\n \u003cli\u003eBisani, M., \u0026amp; Ney, H. (2005). Open vocabulary speech recognition with flat hybrid models. In INTERSPEECH (pp. 725\u0026ndash;728).\u003c/li\u003e\n \u003cli\u003eBourlard, H. A., \u0026amp; Morgan, N. (1993). Connectionist Speech Recognition: A Hybrid Approach. Kluwer Academic Publishers, Norwell, MA, USA. ISBN 0792393961.\u003c/li\u003e\n \u003cli\u003eDavis, S., \u0026amp; Mermelstein, P. (1980). Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. IEEE Transactions on Acoustics, Speech and Signal Processing, 28(4), 357\u0026ndash;366.\u003c/li\u003e\n \u003cli\u003eBaye, A., Tachbelie, Y., \u0026amp; Besacier, L. (2021). End-to-end Amharic speech recognition using LAS architecture. In Proceedings of the 12th Language Resources and Evaluation Conference (pp. 4058-4065). European Language Resources Association.\u003c/li\u003e\n \u003cli\u003eDumitru, C. O., \u0026amp; Gavat, I. (2006, June). A Comparative Study of Feature Extraction Methods Applied to Continuous Speech Recognition in Romanian Language. In Multimedia Signal Processing and Communications, 48th International Symposium ELMAR-2006 (pp. 115-118). IEEE.\u003c/li\u003e\n \u003cli\u003eFurui, S., Ichiba, T., Shinozaki, T., Whittaker, E. W., \u0026amp; Iwano, K. (2005). Cluster-based modeling for ubiquitous speech recognition. Interspeech 2005, 2865-2868.\u003c/li\u003e\n \u003cli\u003eGalescu, L. (2003). Recognition of out-of-vocabulary words with sub-lexical language models. In INTERSPEECH.\u003c/li\u003e\n \u003cli\u003eGebremedhin, Y. B., Duckhorn, F., Hoffmann, R., \u0026amp; Kraljevski, I. (2013). A new approach to develop a syllable-based, continuous Amharic speech recognizer. July, 1684\u0026ndash;1689.\u003c/li\u003e\n \u003cli\u003eGraves, A. (2012). Supervised Sequence Labelling with Recurrent Neural Networks, volume 385 of Studies in Computational Intelligence. Springer.\u003c/li\u003e\n \u003cli\u003eHebash H.O. Nasereddin, A. A. (January 2018). Classification techniques for automatic speech recognition (ASR) algorithms used with real-time speech translation. 2017 Computing Conference. London, UK: IEEE.\u003c/li\u003e\n \u003cli\u003eHinton, G., Deng, L., Yu, D., Dahl, G., Mohamed, A. r., Jaitly, N., Senior, A., Vanhoucke, V., Nguyen, P., Sainath, T., \u0026amp; Kingsbury, B. (2012). Deep neural networks for acoustic modeling in speech recognition. Signal Processing Magazine.\u003c/li\u003e\n \u003cli\u003eJaitly, N., \u0026amp; Hinton, G. E. (2011). Learning a better representation of speech soundwaves using restricted Boltzmann machines. In ICASSP (pp. 5884\u0026ndash;5887).\u003c/li\u003e\n \u003cli\u003ePokhariya, J. S., \u0026amp; D. S. (2014). Sanskrit Speech Recognition using Hidden Markov Model Toolkit. International Journal of Engineering Research \u0026amp; Technology, 3(10).\u003c/li\u003e\n \u003cli\u003eKebebew, T. (2010). Speaker-dependent speech recognition for Afan Oromo using hybrid hidden Markov models and artificial neural network. Addis Ababa, Ethiopia: Addis Ababa University.\u003c/li\u003e\n \u003cli\u003eSolomon Teferra Abate, W. M. (n.d.). Automatic Speech Recognition for an UnderResourced Language \u0026ndash; Amharic. Department of Informatics, Natural Language Systems Group, University of Hamburg, Germany.\u003c/li\u003e\n \u003cli\u003eTeferra, S., \u0026amp; Menzel, W. (2007). Syllable-based speech recognition for Amharic. June, 33\u0026ndash;40.\u003c/li\u003e\n \u003cli\u003eTohye, T. G. (2015). Towards Improving the Performance of Spontaneous Amharic Speech Recognition. Master\u0026apos;s thesis. Addis Ababa University, Addis Ababa, Ethiopia.\u003c/li\u003e\n \u003cli\u003eYifru, M. (2003). Application of Amharic Speech Recognition System to Command. Master\u0026apos;s thesis. Addis Ababa University, Addis Ababa, Ethiopia.\u003c/li\u003e\n \u003cli\u003eGraves, A., Fern\u0026aacute;ndez, S., Gomez, F., \u0026amp; Schmidhuber, J. (2006). Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks. In ICML, pages 369\u0026ndash;376. ACM.\u003c/li\u003e\n \u003cli\u003eJurafsky, D., \u0026amp; Martin, J. H. (2009). Speech and Language Processing (2nd ed.). Prentice Hall.\u003c/li\u003e\n \u003cli\u003eHuang, X., Acero, A., \u0026amp; Hon, H. W. (2001). Spoken Language Processing: A Guide to Theory, Algorithm, and System Development. Prentice Hall PTR.\u003c/li\u003e\n \u003cli\u003eGraves, A., Mohamed, A.-r., \u0026amp; Hinton, G. (2013). Speech Recognition with Deep Recurrent Neural Networks. In IEEE International Conference on Acoustics, Speech and Signal Processing (pp. 6645\u0026ndash;6649).\u003c/li\u003e\n \u003cli\u003eHinton, G., Deng, L., Yu, D., Dahl, G. E., Mohamed, A.-r., Jaitly, N., Senior, A., Vanhoucke, V., Nguyen, P., Sainath, T. N., \u0026amp; Kingsbury, B. (2012). Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups. IEEE Signal Processing Magazine, 29(6), 82\u0026ndash;97.\u003c/li\u003e\n \u003cli\u003eGraves, A., \u0026amp; Jaitly, N. (2014). Towards End-to-End Speech Recognition with Recurrent Neural Networks. In International Conference on Machine Learning (pp. 1764\u0026ndash;1772).\u003c/li\u003e\n \u003cli\u003eJelinek, F., \u0026amp; Mercer, R. L. (1980). Interpolated Estimation of Markov Source Parameters from Sparse Data. In Proceedings of the Workshop on Pattern Recognition in Practice (pp. 381\u0026ndash;397).\u003c/li\u003e\n \u003cli\u003eDeng, L., \u0026amp; Yu, D. (2014). Deep Learning: Methods and Applications. Foundations and Trends in Signal Processing.\u003c/li\u003e\n \u003cli\u003eYoung, S., Evermann, G., Gales, M., Hain, T., Kershaw, D., Liu, X., Moore, G., Odell, J., Ollason, D., Povey, D., Valtchev, V., \u0026amp; Woodland, P. (2006). The HTK Book (for HTK Version 3.4). Cambridge University Engineering Department.\u003c/li\u003e\n \u003cli\u003eAbdel-Hamid, O., \u0026amp; Jiang, H. (2013). Fast speaker adaptation of deep neural networks using model compression. In Proceedings of the 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (pp. 7822-7826). IEEE.\u003c/li\u003e\n \u003cli\u003eChavan, S., \u0026amp; Sable, P. (2013). Speech recognition using hidden Markov model. International Journal of Engineering Research and Applications, 3(4), 2319-2324.\u003c/li\u003e\n \u003cli\u003eNovoa, J., Lleida, E., \u0026amp; Hernando, J. (2018). A comparative study of HMM-based and DNN-based ASR systems for under-resourced languages. Computer Speech \u0026amp; Language, 47, 1-22.\u003c/li\u003e\n \u003cli\u003ePalaz, D., Kılı\u0026ccedil;, R., \u0026amp; Yılmaz, E. (2019). A comparative study of deep neural network and hidden Markov model based acoustic models for Turkish speech recognition. Journal of King Saud University-Computer and Information Sciences.\u003c/li\u003e\n \u003cli\u003eLee, K., \u0026amp; Kim, H. (2019). Dynamic Wrapping for Robust Speech Recognition in Adverse Environments. IEEE Access, 7, 129541-129550.\u003c/li\u003e\n \u003cli\u003eWang, Y., \u0026amp; Wang, D. (2019). A novel dynamic wrapping method for robust speech recognition under noisy environments. Journal of Ambient Intelligence and Humanized Computing, 10(5), 1825-1834.\u003c/li\u003e\n \u003cli\u003eChorowski, J. K., Bahdanau, D., Serdyuk, D., Cho, K., \u0026amp; Bengio, Y. (2015). Attention-based models for speech recognition. In Advances in Neural Information Processing Systems (pp. 577-585).\u003c/li\u003e\n \u003cli\u003eKim, J., Park, S., Kang, K., \u0026amp; Lee, H. (2017). Joint CTC-attention based end-to-end speech recognition using multi-task learning.\u003c/li\u003e\n \u003cli\u003eZhang, Y., Deng, K., Cao, S., \u0026amp; Ma, L. (2021). Improving Hybrid CTC/Attention End-to-end Speech Recognition with Pretrained Acoustic and Language Model.\u003c/li\u003e\n \u003cli\u003eDeng, K., Zhang, Y., Cao, S., \u0026amp; Ma, L. (2022). Improving CTC-based speech recognition via knowledge transferring from pre-trained language models.\u003c/li\u003e\n \u003cli\u003eChen, C. (2021). ASR Inference with CTC Decoder. PyTorch.\u003c/li\u003e\n \u003cli\u003eHannun, A., Case, C., Casper, J., Catanzaro, B., Diamos, G., Elsen, E., \u0026hellip; \u0026amp; Prenger, R. (2014). Deep speech: Scaling up end-to-end speech recognition.\u003c/li\u003e\n \u003cli\u003eIgor Macedo Quintanilha (2017). End-to-end speech recognition applied to Brazilian Portuguese using deep learning.\u003c/li\u003e\n \u003cli\u003eHanan Aldarmaki, Asad Ullah, and Nazar Zaki (2021). Unsupervised Automatic Speech Recognition: A Review. arXiv.org, 2021, arxiv.org/abs/2106.04897.\u003c/li\u003e\n \u003cli\u003eAbdelrahma Ahmed, Yasser Hifny, Khaled Shaalan, and Sergio Tolan (2016). Lexicon Free Arabic Speech Recognition. link.springer.com/chapter/10.1007/978-3-319-48308-5_15.\u003c/li\u003e\n \u003cli\u003eWren, Y., Titterington, J., \u0026amp; White, P. (2021). How many words make a sample? Determining the minimum number of word tokens needed in connected speech samples for child speech assessment. Clinical Linguistics \u0026amp; Phonetics, 35(8), 761-778. https://doi.org/10.1080/02699206.2020.1827458.\u003c/li\u003e\n \u003cli\u003eMalik, M., Malik, M. K., Mehmood, K., \u0026amp; Makhdoom, I. (2021). Automatic speech recognition: a survey. Multimedia Tools and Applications, 80(6), 9411-9457. https://doi.org/10.1007/s11042-020-10073-7.\u003c/li\u003e\n \u003cli\u003eZhang, X., \u0026amp; Wang, H. (2011). A syllable-based connected speech recognition system for Mandarin. In 2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 4856-4859). IEEE. https://doi.org/10.1109/ICASSP.2011.5947265.\u003c/li\u003e\n \u003cli\u003eKhan, A. I., \u0026amp; Zahid, S. (2012). A comparative study of speech recognition techniques for Urdu language. International Journal of Speech Technology, 15(4), 497-504. http://doi.org/10.1007/s10772-012-9175-z.\u003c/li\u003e\n \u003cli\u003eChen, C.-H., \u0026amp; Chen, C.-W. (2010). A novel speech recognition system based on a self-organizing feature map and a fuzzy expert system. Microprocessors and Microsystems, 34(6), 242-251. http://doi.org/10.1016/j.micpro.2010.04.003\u003c/li\u003e\n \u003cli\u003eGe, X., Wu, L., Xia, D., \u0026amp; Zhang, P. (2013). A multi-objective HSA of nonlinear system modeling for hot skip-passing. In 2013 Sixth International Conference on Advanced Computational Intelligence (ICACI) (pp. 1-6). IEEE. https://doi.org/10.1109/ICACI.2013.6748484\u003c/li\u003e\n \u003cli\u003eVilla, A. E. P., Masulli, P., \u0026amp; Pons Rivero, A. J. (Eds.). (2016). Artificial neural networks and machine learning \u0026ndash; ICANN 2016: 25th International Conference on Artificial Neural Networks, Barcelona, Spain, September 6-9, 2016, Proceedings, Part I. Springer International Publishing. https://doi.org/10.1007/978-3-319-44781-0\u003c/li\u003e\n \u003cli\u003eChen, X., \u0026amp; Hu, Y. (2012). Learning optimal warping window size of DTW for time series classification. In 2012 11th International Conference on Information Science, Signal Processing, and their Applications (ISSPA) (pp. 1-4). IEEE. https://doi.org/10.1109/ISSPA.2012.6310488\u003c/li\u003e\n \u003cli\u003eLa Rosa, M., Loos, P., \u0026amp; Pastor, O. (Eds.). (2016). Business Process Management: 14th International Conference, BPM 2016, Rio de Janeiro, Brazil, September 18-22, 2016. Proceedings. Springer. https://doi.org/10.1007/978-3-319-45348-4\u003c/li\u003e\n \u003cli\u003eNguyen, V. N. (2016). A framework for business process improvement: A case study of a Norwegian manufacturing company (Master\u0026rsquo;s thesis). University of Stavanger. https://brage.bibsys.no/xmlui/bitstream/id/430871/16-00685-3%20Van%20Nhan%20Nguyen%20-%20Master\u0026rsquo;s%20Thesis.pdf%20266157_1_1.pdf\u003c/li\u003e\n \u003cli\u003eLeCun, Y., Bottou, L., Bengio, Y., \u0026amp; Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. https://doi.org/10.1109/5.880079\u003c/li\u003e\n \u003cli\u003eIentilucci, E. J. (2006). Statistical models for physically derived target sub-spaces. In S. S. Shen \u0026amp; P. E. Lewis (Eds.), Imaging Spectrometry XI (Vol. 6302, p. 63020A). SPIE. https://doi.org/10.1117/12.679525\u003c/li\u003e\n \u003cli\u003eBegel, A., \u0026amp; Bosch, J. (2013). The DevOps phenomenon. Communications of the ACM, 56(11), 44-49. http://dl.acm.org/citation.cfm?id=2113113.2113141\u003c/li\u003e\n \u003cli\u003eGoodfellow, I., Bengio, Y., \u0026amp; Courville, A. (2016). Deep Learning. MIT Press.\u003c/li\u003e\n \u003cli\u003eScherzer, O. (Ed.). (2015). Handbook of mathematical methods in imaging (2nd ed.). Springer. https://doi.org/10.1007/978-3-319-46478-7\u003c/li\u003e\n \u003cli\u003ePersagen Consulting. (n.d.). Machine learning. https://persagen.com/files/ml.html\u003c/li\u003e\n \u003cli\u003eSaha, S. K., \u0026amp; Tripathy, A. K. (2017). Automated IT system failure prediction: A deep learning approach. International Journal of Computer Applications, 159(9), 1\u0026ndash;6. https://www.researchgate.net/publication/313456329_Automated_IT_system_failure_prediction_A_deep_learning_approach\u003c/li\u003e\n \u003cli\u003eKotera, J., \u0026Scaron;roubek, F., \u0026amp; Milanfar, P. (2013). Blind deconvolution using alternating maximum a posteriori estimation with heavy-tailed priors. In A. Petrosino, L. Maddalena, \u0026amp; P. Soda (Eds.), Computer analysis of images and patterns (pp. 59\u0026ndash;66). Springer. https://doi.org/10.1007/978-3-642-40246-3_8\u003c/li\u003e\n \u003cli\u003eArtificial intelligence. (2022, January 5). In Wikipedia. https://en.wikipedia.org/wiki/Artificial_intelligence\u003c/li\u003e\n \u003cli\u003eMwiti, D. (2019, September 4). A 2019 guide for automatic speech recognition. Heartbeat. https://heartbeat.fritz.ai/a-2019-guide-for-automatic-speech-recognition-f1e1129a141c\u003c/li\u003e\n \u003cli\u003ePersagen. (n.d.). Machine learning. Retrieved January 6, 2022, from https://persagen.com/files/ml.html\u003c/li\u003e\n \u003cli\u003eKlein, A., \u0026amp; Kienle, A. (2012). Supporting distributed software development by modes of collaboration. Multimedia Systems, 18(6), 509\u0026ndash;520. http://doi.org/10.1007/s00530-012-0266-0\u003c/li\u003e\n \u003cli\u003eChen, Y., \u0026amp; Cong, J. (2017). DAPlace: Data-aware three-dimensional placement for large-scale heterogeneous FPGAs. In 2017 IEEE/ACM International Conference on Computer-Aided Design (ICCAD) (pp. 236\u0026ndash;243). IEEE. http://doi.org/10.1109/ICCAD.2017.8203812\u003c/li\u003e\n \u003cli\u003eZhang, Y., Zhang, H., Lin, H., \u0026amp; Zhang, Y. (2017). A new method for remote sensing image registration based on SURF and GMS. Remote Sensing, 9(3), 298. http://doi.org/10.3390/rs9030298\u003c/li\u003e\n \u003cli\u003eSchmidhuber, J. (2000). How to count time: A simple and general neural mechanism. ftp://ftp.idsia.ch/pub/juergen/TimeCount-IJCNN2000.pdf\u003c/li\u003e\n \u003cli\u003eGoodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., \u0026amp; Bengio, Y. (2014). Generative adversarial nets. https://arxiv.org/pdf/1406.1078v3.pdf\u003c/li\u003e\n \u003cli\u003eSrivastava, A. (2018, November 14). Basic architecture of RNN and LSTM. PyDeepLearning. https://pydeeplearning.weebly.com/blog/basic-architecture-of-rnn-and-lstm\u003c/li\u003e\n \u003cli\u003eGraves, A., Mohamed, A.-R., \u0026amp; Hinton, G. (2013). Speech recognition with deep recurrent neural networks. 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 6645\u0026ndash;6649. https://www.cs.toronto.edu/~graves/asru_2013.pdf\u003c/li\u003e\n \u003cli\u003eZhang, Y., Zhang, J., Zhang, J., \u0026amp; Li, H. (2020). A survey on deep learning for named entity recognition. IEEE Transactions on Knowledge and Data Engineering, 32(9), 1635\u0026ndash;1658. https://doi.org/10.1109/TKDE.2019.2906425\u003c/li\u003e\n \u003cli\u003eShaik Johny Basha; Duggineni Veeraiah; Boddu Venkat Charan; Wiltrud Sahithi Yeddu; Devalla Ganesh Babu (2023)\u003c/li\u003e\n \u003cli\u003eOlah, C. (2015, August 27). Understanding LSTM networks. http://colah.github.io/posts/2015-08-Understanding-LSTMs/\u003c/li\u003e\n \u003cli\u003eEjigu, Y.A.; Asfaw, T.T. Enhancing Amharic Speech Recognition in Noisy Conditions through End-to-End Deep Learning. Preprints 2024, 2024020754. https://doi.org/10.20944/preprints202402.0754.v1\u003c/li\u003e\n \u003cli\u003eEjigu, Y.A.; Asfaw, T.T. Large Scale Speech Recognition for Low Resource Language Amharic, an End-to-End Approach. Preprints 2024, 2024020813. https://doi.org/10.20944/preprints202402.0813.v1\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"international-journal-of-speech-technology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ijst","sideBox":"Learn more about [International Journal of Speech Technology](http://link.springer.com/journal/10772)","snPcode":"10772","submissionUrl":"https://submission.springernature.com/new-submission/10772/3","title":"International Journal of Speech Technology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Amharic Speech Recognition, End-to-End Deep Learning, Noise-Robust ASR, BiGRU, Connectionist Temporal Classification, Real-World Noisy Environments, Automatic Feature Extraction, Low-Resource Languages, Spectral Subtraction, Subspace Filtering","lastPublishedDoi":"10.21203/rs.3.rs-6419413/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6419413/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEnd-to-end automatic speech recognition (ASR) offers a streamlined alternative to traditional systems that rely on multiple, separately trained language, acoustic, and pronunciation models. In this paper, we present a noise-robust, end-to-end ASR framework tailored specifically to the Amharic language. Our approach integrates a convolutional neural network (CNN), a recurrent neural network (RNN), and Connectionist Temporal Classification (CTC) to directly transcribe speech into text\u0026mdash;bypassing the need for labor-intensive dictionary creation. We evaluate our method on a large corpus of 20,000 noisy Amharic utterances, achieving a word error rate (WER) of just 7%. This result underlines the effectiveness of our system in handling challenging acoustic conditions. By reducing complexity and manual overhead, our end-to-end model offers a practical and accurate solution for real-world deployments, with broader implications for developing ASR in other low-resource and noise-prone environments.\u003c/p\u003e","manuscriptTitle":"A Noise-Robust End-to-End Framework for Amharic Speech Recognition","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-04 11:03:06","doi":"10.21203/rs.3.rs-6419413/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-10-18T21:46:24+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-02T12:25:14+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"231267520566917227189906225611313762352","date":"2025-08-07T11:29:28+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-29T15:01:35+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-19T11:22:17+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-19T11:19:49+00:00","index":"","fulltext":""},{"type":"submitted","content":"International Journal of Speech Technology","date":"2025-04-10T10:49:55+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"international-journal-of-speech-technology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ijst","sideBox":"Learn more about [International Journal of Speech Technology](http://link.springer.com/journal/10772)","snPcode":"10772","submissionUrl":"https://submission.springernature.com/new-submission/10772/3","title":"International Journal of Speech Technology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"29b1585e-943f-4eae-88a3-c622b7ad0834","owner":[],"postedDate":"August 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-12-29T16:05:37+00:00","versionOfRecord":{"articleIdentity":"rs-6419413","link":"https://doi.org/10.1007/s10772-025-10224-x","journal":{"identity":"international-journal-of-speech-technology","isVorOnly":false,"title":"International Journal of Speech Technology"},"publishedOn":"2025-12-24 15:57:59","publishedOnDateReadable":"December 24th, 2025"},"versionCreatedAt":"2025-08-04 11:03:06","video":"","vorDoi":"10.1007/s10772-025-10224-x","vorDoiUrl":"https://doi.org/10.1007/s10772-025-10224-x","workflowStages":[]},"version":"v1","identity":"rs-6419413","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6419413","identity":"rs-6419413","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}