Machine Learning Models of Sounds to Predict Textural Attributes of Snack Chips

Machine Learning Models of Sounds to Predict Textural Attributes of Snack Chips | 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 Machine Learning Models of Sounds to Predict Textural Attributes of Snack Chips Toby Serrano, William Kerr This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6514631/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Crispness, crunchiness and crackliness attributes of four snack chips (Lays Classic, Kettle-Cooked, Ruffles and Doritos) was assessed through mechanical and acoustic signatures, coupled with sensory analysis by a trained panel. The highest crispness levels were found for Lay’s Classic chips and greatest crunchiness for Kettle Cooked chips. However, any chip had simultaneous perceptions of crispness, crunchiness and crackliness. For mechanical properties, crunchiness and crackliness were most closely associated with peak force and “linear distance”; for sound they were most closely related to sound intensity and the amount of acoustical energy below 1.7 kHz. A machine learning (ML) model was built based on 200 sound files from each chip, augmented with noise to supply 1000 files for each type. The ML model considered two additional categories of “noise” and “unknown” sounds, and used 80% of the datasets for training and 20% for testing. The prediction accuracy ranged from 89.5% for Doritos to 98% for Kettle-Cooked chips. In addition, the accuracy ranged from 90.0-97.0% for predicting crispness, 91.0-97.0% for crunchiness and 92.5-99.0% for crackliness. Crispness crunchiness AI models acoustic measurements Figures Figure 1 Figure 2 Introduction Texture is an important attribute of many food products, directly influencing consumer preferences and marketability (Szczesniak and Kahn 1971). However, there are specific texture properties that present as foods break abruptly into several pieces when chewed and are accompanied by complex sounds. Several terms are used to describe these attributes, and they may be used interchangeably by consumers. “Crispness” is one of these and regarded as one of the most desired attributes in food. Szczesniak and Kahn (1984) found that consumers view crispness as well-liked, a sign of excellent cooking, and “very synonymous with freshness and wholesomeness”. They also showed that crispness was versatile, in that it could be paired with other texture attributes such as softness, smoothness, and creaminess. “Crunchiness” is a related textural attribute enjoyed nearly as much as crispness. In their research Szczesniak & Kahn (1984) indicated that crunchiness adds contrast and liveliness to food texture. As with crispness, crunchiness is desirable in combination with many other textures. The public perception of crunchiness is highly positive and “regarded with warmth and affection”. “Crackliness” is another textural attribute found in snack foods but has not been as well studied as crispness or crunchiness. Crackliness can be defined as “multiple sharp repeated noises when a food is bitten and chewed” (Vickers 1984a). Manufacturers and researchers have struggled to find methods to determine the quality and intensity of crisp, crunchy or crackly texture attributes. Of course, the primary means of assessment is through sensory analyses performed by human subjects. This may involve untrained panelists that are tasked to determine the level of likability or preference of one product over another. In contrast, trained or experienced persons can be used to determine the intensity level of one or more attributes such as crunchiness, hardness or chewability. Instrumental texture analysis is valuable for quickly assessing changes in quality, comparing products or determining the effects of new formulations (Roudaut et al. 2002). This typically involves compressive, shear or puncture probes that are lowered through the sample, and the force-distance response measured as the sample deforms and breaks. The crispness of many different types of foods has been analyzed using such mechanical methods. This includes bread crust (Primo-Martín et al. 2008), potato chips (Salvador et al. 2009), wafers (Çarşanba et al. 2018), crisp bread (Aleixandre et al. 2021), almonds (Cheely et al. 2018) and extruded cereals (Chanvrier et al. 2014), as well as dried fruits such as mango slices (Link et al. 2018) and bananas (de Moraes et al. 2022). One challenge is how to best analyze rather noisy data to provide a metric that reflects crispness and similar attributes. For simple systems such as saltine crackers or almonds (Cheely et al. 2018), this may be a single maximum or “snap” force that correlates with crispness. Likewise, the force or distance to the first fracture event may be deemed important. Many approaches take note of the numerous fracture events that occur under stress. Thus, Prabhakar et al. (2023) found the ratio of the initial fracture force to peak force useful in assessing pecan crunchiness. Others have enumerated the total number of peaks during breakage. Suwonsichon and Peleg (1998) studied fluctuations of the force-displacement relationship, as quantified by its fractal dimension, which was observed to change with varying moisture contents in the cereals. Another approach is to measure the deviations by summing the total distances for each point-to-point length, as was done for corn chips (Xu and Kerr 2012). If force data is collected over time such that there are n data points, the “linear distance” is given by: Crisp, crunchy and crackly foods generally make sounds as they are consumed. Dacremont (1995) recorded both air and bone conducted sounds as panelists chewed a variety of crisp, crunchy or crackly foods. Chen et al. (2005) described an “Acoustic Envelope Detector” attached to a texture analyzer for concurrently obtaining acoustic and force data from food samples. For biscuits, it was found that for every force drop, there was a simultaneous auditory occurrence. When studying the effects of water activity Castro-Prada et al. (2009) found that greater sound intensity related to a higher level of “crispness”. Multiple acoustic parameters have been calculated to assess textural properties. In research on wafers (Çarşanba et al. 2018), potato chips (Salvador et al. 2009; Taniwaki and Kohyama 2012) and almonds (Varela et al. 2006), greater maximum sound pressure (S max ) and number of acoustic peaks in the audio waveform were positively correlated with sensory crispness. When studied simultaneously, it has been noted that the fluctuations in both mechanical and acoustical curves are an indication of crispness (Salvador et al. 2009; Link et al. 2018; Chen et al. 2005; Vincent 1998). Research has also included frequency analysis, particularly when comparing crispy to crunchy foods. Dacremont (1995) studied crispy food samples including flatbreads, while the crunchy samples included foods such as raw carrots. They found that crisp foods generate more high frequency sounds (>5kHz), crunchy foods have more sound energy at <1.7 kHz and crackly foods generate low pitched sounds with a high level of bone conduction. A promising approach to categorizing sounds is through machine learning (ML). This consists of recording relevant audio events, processing these to generate useful features while reducing data size (Lyons 2009), then using these to train models using large datasets. Training consists of analyzing the data with neural network algorithms, recognizing the data patterns, and using the resulting set of nested equations to identify unknown items. For example, Nimbarte et al. (2023) used Python to build an ML model to detect crying babies, with an accuracy of 0.8 and precision of 0.48. An ML model was developed with Google Colab to detect COVID-19 cases from the audio spectrograms of coughing, sneezing, and respiratory sounds (Rodriguez et al. 2020). ML has also been used to recognize various animal species (Balemarthy et al. 2018) and birds (Das et al. 2020) based on their vocalizations. Trivedi and Shroff (2021) utilized an Arduino microcomputer and Edge Impulse to classify the species of deadly mosquitoes based on recorded audio data of mosquito wing beats. For food research, Kato et al. (2019) developed an ML-based system to estimate crispness from force and sound data. However, there was no inputs of human perceptions of crispness. In this research, an ML model was trained to classify the sounds of different snack chips, with the goal that after training, the model should be able to determine the kind of snack food based on the sound it emits during compression. Concurrently, the snack foods were evaluated for their textural attributes using sensory assessments, force-time response, and traditional acoustic calculations, to provide greater understanding of the causes of any differences. In addition, ML models were constructed to differentiate and predict levels of crispness, crunchiness or crackliness based on sensory evaluations of each of these attributes. Materials and methods Samples The chips used were Lay's ® Classic Potato Chips, Lay's ® Kettle Cooked Original Potato Chips, Ruffles ® Original Potato Chips, and Doritos ® Nacho Cheese Flavored Tortilla Chips (Frito-Lay, Plano, Texas), purchased from a local grocery store in Athens, Georgia. The chip samples had minimal preparation before analysis and were analyzed immediately after removal from the original packaging to retain freshness. Sensory analysis All samples were evaluated by trained panel using quantitative descriptive analysis (Chauvin et al. 2008; Stone et al. 2020). Nine participants (3M, 6F; age range 20 to 60 y) were recruited from the University of Georgia and operated under protocols approved by the UGA Institutional Review Board (PROJECT00006724). Panelists received a small compensation for their participation in the study. The sensory analysis was conducted through six different sessions, each held on a different day. In the first session, panelists were screened for hearing, dental, or other issues that would prevent consumption of hard snack foods. Panelists were taught to sample in a specific manner as follows: hold one piece of the sample, take one bite of the sample using the incisor teeth, and then chew using the molar teeth and swallow; after each bite, rinse the mouth with filtered water, then expectorate the rinse water into a provided container. During training, participants discussed textural attributes that best described all the samples and concluded that the attributes “crispness”, “crunchiness”, and “crackliness” were the essential textural attributes of the chip snacks. Panelists were then provided definitions of “crispness”, “crunchiness”, and “crackliness” as described in Chauvin et al. (2008). The group next developed a new standard scale for each of the textural attributes. The foods were prepared by placing each sample in a 2-oz lidded plastic cup. In the next session, panelists were asked to rate each food standard on a 15-point scale, with 15 being the most extreme level of the textural attribute, and 0 being a complete lack of that attribute. Panelists rated each food first by themselves, writing the name of each food on a point of the line. Panelists afterward discussed among themselves to agree in unanimity on where each food belongs on the 15-point line. In subsequent sessions, panelists were asked to rate the four different chip samples on the three standard scales for crispness, crunchiness, and crackliness. At this stage, evaluations occurred in isolated sensory booths. Each of the different chip types were labeled with a randomized three-digit code. Panelists were provided with a sample of each food, a napkin, rinsing water, an expectorant container, a pen, and a paper with the three developed standard scales. Standards were also provided so panelists could remind themselves of the standard scale. In the fourth, fifth, and sixth sessions, panelists again rated the four different kinds of chips. Instrumental texture analyses The four chip samples were analyzed using a TA.XT Plus texture analyzer (Stable Micro Systems, Surrey, UK) fitted with a “Crisp Fracture Support Rig”. Chip samples were placed on a cylindrical stand 25 mm in diameter as a 6.2 mm ball tip probe descended through the center of the sample at 5 mm/s. The force versus time data was analyzed using the software, determining the peak force in Newtons (N), the linear distance (AU), the area under the curve in Ns, and the number of peaks in the curve using a threshold of 68.6 mN. At least 200 replications of chps of each type were analyzed. Acoustic analysis An Arduino Nano 33 BLE Sense microcomputer connected to a laptop was positioned 10 mm away from the chip sample. The Arduino’s PDM microphone recorded the sounds during the crushing and puncture of the chip sample by the probe. At least 200 unique audio files were recorded for each of the chip types. Using an in-house MATLAB (R2023b) script, the audio files were analyzed for their sound intensity, number of sound peaks, and percentage of cumulative frequencies under 1.7 kHz (Mathwork Inc, Nattick, MA). Machine learning models The “.wav” files were trimmed to 5 s each and each was given a label for its corresponding type of chip. At least 200 unique audio files were recorded for each chip type. To prevent overfitting, the sets of audio files were processed in Python 3.10 using the Google Speech Commands dataset (Hymel 2022). This data augmentation tool mixed the chip audio files with background noise to create more audio files for each chip type. With this procedure, 800 augmented chip audio files were created, totaling 1000 chip audio files for each chip type: 4000 in total. In addition, audio samples of sound representing “unknown” sound and “noise” were obtained from Shawn Hymel’s GitHub programming code repository, part of Edge Impulse’s Coursera online learning platform (Hymel 2022). For both “unknown” and “noise”, 1000 audio files were obtained: 2000 in total. The 6000 audio files comprising direct chip sounds, ‘noise’ and ‘unknown’ sounds (totaling 8 h and 20 min) of audio, were imported into the Edge Impulse platform. The entire dataset was randomly split into one training and one test set. The training set encompassed 80 percent of the data (4800 files), while the test set encompassed 20 percent of the data (1200 files). The files were then edited in the processing block. The ‘window size’, that is the amount of data to be converted for each classification, was set to the length of each audio file, that is 5000 ms. The ‘window increase’ (used to artificially create more features) was 500 ms, the frequency of the data was 16000 Hz, and zero-pad setting was enabled. The Mel frequency cepstral coefficients technique was used for feature extraction. The MFCCs represents a reduced dataset that captures critical temporal and frequency information of the audio signal that is most relevant to the model. Research has shown that MFCCs are effective in classifying sounds as they consider sound energy in frequency bands that mimic the human auditory response (Abdul and Al-Talabani 2022). The parameters used for the MFCC design were as follows. The number of cepstral coefficients was 13; the length of each frame was 20 ms; the frame stride was 20 ms; the number of filters in the filter bank was 32; the FFT length was 256; the normalization window size was 101; and the lowest frequency was 0 Hz. The high frequency was set to half of the sample rate and the pre-emphasizing coefficient was 0.98. For neural network training, the number of training cycles was 100 and the learning rate set to 0.005. The validation set was 20% of the training set. The Profile int8 model setting was turned on. This is a quantization tool that converts data to 8-bit integers instead of floating-point numbers and integer math instead of floating-point math, thus reducing memory and computing requirements. The first layer was an Input layer of 3250 features. The second layer was a Reshape layer made up of 13 columns. The third layer was a 1-dimensional single convolution layer followed by a pooling layer, made of 8 neurons and a kernel size of 3. The fourth layer was a Dropout layer, with a dropout rate of 0.25. The fifth layer was a 1-dimensional single convolution/pool layer of 25 neurons and a kernel size of 25. The sixth layer was another Dropout layer with a 0.25 dropout rate. The seventh layer was a Flatten layer. The eighth and final layer was the output layer, with the 6 different classes of Lays Classic, Kettle, Ruffles, Doritos, “unknown”, and “noise”. Statistical analyses The data were initially analyzed using the Shapiro-Wilks test (SPSS 29 Statistics software, Armank, NY). The test is based on the null hypothesis that the data set is normally distributed and outputs a value from 0 to 1, where 1 indicates the data closely resembles the normal distribution. Thus, p <0.05 signifies that the assumption of normality must be rejected. In concert, data were tested for homogeneity of variance using Levene’s test. Here, p<0.05 indicates that the null hypothesis of equal variances is rejected and there is a difference between the variances in the population. When normality and equal variances could not be assured, the Kruskal-Wallis was used to test whether samples originated from the same distribution. Thus, it serves as the non-parametric equivalent of the one-way ANOVA. When differences were found, the post-hoc Dunn’s test was used as a pairwise comparison test that assesses the differences between group medians. Principal component analysis was done using JMP 12 (JMP Statistical Discovery LLC, Cary, NC). Results and discussion Sensory analysis Results from the modified quantitative descriptive analyses of the chips are shown in Figure 1. The highest crispness intensity (11.5) was found for the Lay’s Classic potato chips and lowest for the Doritos tortilla chips (8.14). The greatest crunchiness (11.6) and crackliness (10.4) was noted for the Lay’s Kettle Cooked potato chips. For Lay’s Classic chips the crispness (at 11.5) was clearly the predominant characteristic with scores of 6.35 and 6.5 for crunchiness and crackliness, respectively. Some researchers have suggested that foods perceived as more crisp than crunchy are less dense (Tunick et al. 2013), and this may be why the Lay’s Classic chips were perceived as more crisp than crunchy. For the other chips, the values of the texture attributes were more similar. Thus, for the Lay’s Kettle Cooked chips values of crispness, crunchiness and crackliness were 10.5, 11.6 and 10.4, while for the Ruffles, these values were 9.29, 9.36 and 9.16. This shows that these textural terms are not disconnected, and even trained panelists can perceive more than one attribute for each food item. However, for some foods one of the textural percepts may predominate. With this understanding, some researchers have even used sensory ballots with terminology such as “more crisp than crunchy” or “more crunchy than crisp” (b 1984). The Shapiro-Wilk test showed p<0.05 for all crispness evaluations, as well as for crunchiness and crackliness for the Lay’s Classic and Kettle Cooked. Thus, most data could not be said to be normally distributed. Further, Levene’s test showed that only crispness ratings had a p > 0.05 so that only the crispness ratings had homogeneity of variance. This may be traced to the fact that the chips did not have uniform shape and size. Because of the lack of normality and homogeneity of variance, non-parametric tests for significant differences were used. The Kruskal-Wallis test showed that p <0.05 for crispness, crunchiness, and crackliness ratings indicating the likelihood of differences amongst the attributes. The Dunn’s test was subsequently used to test for significant differences between paired groups (Table 1). For crispness, p <0.05 when comparing each chip against another. Thus, each chip had a distinguishable level of crispness compared to the others. For crunchiness, p 0.05 for Kettle-Ruffle, Kettle-Dorito, and Ruffle-Dorito, indicating that crackliness ratings were not very distinct amongst chips. Table 1 Dunn’s test for significant difference in sensory descriptive ratings between snack chip types Crispness Crunchiness Crackliness Chip Type p-value p-value p-value Lays-Kettle <0.01 <0.01 <0.01 Lays-Ruffle <0.01 <0.01 0.01 Lays-Dorito <0.01 <0.01 <0.01 Kettle-Ruffle 0.026 0.012 0.567 Kettle-Dorito <0.01 0.092 1.00 Ruffle-Dorito 0.031 1.00 1.00 The results suggest that the panelists’ understanding of crispness was the strongest, leading to a clear difference in ratings between chip types, while the understanding of crunchiness and crackliness was vaguer. Another possibility of course is that the perceived crunchiness and crackliness between Kettle Cooked, Ruffles and Doritos Chips are very similar, leading to an unclear difference in ratings. Instrumental texture analyses Results from the instrumental texture analyses are shown in Table 2. The Kettle Cooked chips showed the greatest peak force (8.32 N), followed by the Doritos (6.49 N), Ruffles (5.07 N) and Lay’s Classic chips (2.75 N). The Kettle Cooked also had the greatest number of fracture peaks (28.4), followed by Doritos (16.8), Ruffles (16.2) and Lay’s Classic (13.2). Finally, the linear distance decreased in the order: Kettle Cooked (65.5 Ns), Doritos (46.3 Ns), Ruffles (40.2 Ns) and Lay’s Classic (30.1 Ns). Thus, the order of decrease was the same for each force-time metric derived from the data. As noted, the Lay’s Classic were perceived as having the greatest sensory rating of crispness yet were lowest in terms of force-related attributes. These force-related metrics have been correlated with crispness in other research (Aleixandre et al. 2021; Suwonsichon and Peleg, 1998; Chaunier et al. 2005; Roudaut et al. 1998). In this case, it seems that texture analyzer values would not be the best predictor of crispness when crunchiness or crackliness are considered at the same time. Table 2 Analyses of force-time response during fracturing of snack chips by instrumental texture analyzer. Chip Type Peak Force (N) Linear Distance (s) Area Under Curve (Ns) Number of Peaks Lays Ave 2.75 30.1 3.09 13.2 Std Dev (0.87) (5.72) (2.07) (7.93) Kettle Ave 8.32 65.5 10.8 28.4 Std Dev (3.94) (33.4) (9.03) (16.4) Ruffle Ave 5.07 40.2 4.77 16.2 Std Dev (1.74) (10.9) (3.72) (8.85) Dorito Ave 6.49 46.3 6.22 16.8 Std Dev (1.42) (13.1) (3.34) (7.96) The Shapiro-Wilk and Levene’s tests all had p <0.05, again indicating the data were not normally distributed nor had homogeneity of variance. The Kruskal-Wallis p -values < 0.05 for every parameter, indicating there were differences in mechanical properties amongst the chip types. The Dunn’s test (Table 3) showed p < 0.05, except for the peak force of “Kettle-Dorito” (p=0.116) and number of peaks for “Ruffle-Dorito” (p=0.99). Thus, except for a few comparisons, the force parameter values were significantly different between every chip type. Table 3 Dunn’s test for significant difference in force parameters between specific chips Peak Force Linear Distance Area Under Curve Number of Peaks Chip Type p-value p-value p-value p-value Lays-Kettle <0.01 <0.01 <0.01 <0.01 Lays-Ruffle <0.01 <0.01 <0.01 0.02 Lays-Dorito <0.01 <0.01 <0.01 <0.01 Kettle-Ruffle <0.01 <0.01 <0.01 <0.01 Kettle-Dorito 0.116 <0.01 0.01 <0.01 Ruffle-Dorito <0.01 <0.01 <0.01 1.00 Acoustical analysis Results from the acoustic analyses of sounds recorded during chip breakage are shown in Table 4. In order, the overall sound intensity decreased from Kettle Cooked (9770 au), Doritos (8616 au), Ruffles (8070 au) and Lay’s Classic (7776 au). The number of sound peaks followed the same order ranging from 34.8 for the Kettle Cooked chips to 16.6 for the Lay’s Classic. Interestingly, this also is the same order found for the peak force and number of force peaks. This reinforces the idea that the sound peaks emanate from individual fracture advents while the sound level is tied to the energy released during the mechanical fracture. Table 4 Analyses of sound recorded during fracturing of snack chips Sound Intensity Number of Sound Peaks % < 1.7 kHz Kettle Ave 9,770 34.8 30.9 Std Dev 2,012 7.34 Dorito Ave 8,616 19.4 31.0 Std Dev 1,663 6.50 Ruffle Ave 8,070 17.7 22.3 Std Dev 968 7.22 Lays Ave 7,777 16.6 23.2 Std Dev 929 6.89 The sound recordings were also analyzed by Fourier transform algorithms to determine the power spectrum density, that is the amount of acoustical energy associated with each frequency. Thus, the analyses plot the relative weights of frequency components that make up the audio signal. Some work has suggested that items sensed as more crunchy than crispy have more acoustical energy at lower frequencies, that is the signal contain more lower “pitches” (Dacremont 1995; Vickers 1984). The percentage of the cumulative frequency under 1.7 kHz for each chip type was also calculated (Table 4). This marker was chosen as snack items with frequencies higher than 1.7 kHz were associated with “crispness”, while frequencies lower than 1.7 kHz were associated with “crunchiness” (Dacremont 1995). Kettle Cooked and Dorito chips both had a percent under 1.7 kHz around 31%, while Lay’s Classic and Ruffles Original had a percentage of 23.2 and 22.3, respectively. This at least concurs with the observation that Lay’s Classic chips were perceived as the crispest while Kettle Cooked were seen as most crunchy, as more of the sound energy resides at higher frequencies. The Shapiro-Wilk test showed p >0.05 for every chip when considering sound intensity, indicating a more normal distribution for this data set. For number of sound peaks p < 0.05, except for Doritos, thus suggesting a non-normal distribution of values. The Levene’s test showed p 0.05 for number of peaks, showing that while the sound intensity values might be considered normally distributed, they did not have homogeneity of variance. The Kruskal-Wallis tests indicated there were significant differences between chip types for the sound properties. Dunn’s Test (Table 5) showed that for sound intensity p < 0.05 between all chip comparisons. Furthermore, the number of sound peaks were different between all types except between “Ruffle-Dorito”. Thus except for one comparison, each chip was distinct in their sound intensity and number of sound peaks. Table 5 Dunn’s test for significant difference in sound parameters between chip types Sound Intensity Number of Sound Peaks Chip Types p-value p-value Lays-Kettle <0.01 <0.01 Lays-Ruffle <0.01 <0.01 Lays-Dorito <0.01 0.07 Kettle-Ruffle <0.01 <0.01 Kettle-Dorito <0.01 <0.01 Ruffle-Dorito 0.01 0.135 As shown in Figure 1, Table 2 and Table 4, crispness ratings did not follow the same pattern as the mechanical and acoustic parameters. That is, the crispness ratings, from highest to lowest, were Lay’s Classic, Kettle Cooked, Ruffles, and Doritos. The mechanical and acoustic parameters were ranked Kettle Cooked, Doritos, Ruffles and Lays Original chips. However, the crunchiness ratings had the exact same ordering as the mechanical and acoustic parameters. This was further analyzed through PCA plots (Figure 2). Thus, crunchiness and to a lesser extent crackliness, were most associated with high peak force and linear distance measurements. The linear correlations for crunchiness and peak force were r 2 =0.98 and for linear distance r 2 =0.94. For crackliness, these were r 2 =0.95 and 0.85, respectively. Other metrics were reasonable predictors of crunchiness and crackliness. Thus, crunchiness was linearly related to the number of force peaks (r 2 =0.86) as well as sound intensity (r 2 =0.79). Crunchiness was most associated with greater acoustic energy at low frequencies, measured by %<1.7kHz. It should be noted that many of the force and acoustic measurements are interdependent, thus each might reasonably serve as a proxy for texture measurement. For example, the number of force peaks was correlated with the number of sound peaks (r 2 =0.99) and overall sound intensity (r 2 =0.97). Likewise, the linear force distance was related to number of peaks or sound intensity (r 2 =0.98) An interesting observation is that crispness was negatively correlated with most other physical measurements. The greatest negative correlation was with crackliness (r 2 =-0.64) and crunchiness (r 2 =-0.44). In terms of physical measurements, crispness was most negatively correlated to peak force (r 2 =-0.40) and %<1.7kHz (r 2 =-0.38). The theory that crisp foods are more associated with higher frequency content and crunchy/crackly foods have more low frequencies is borne out by the results. However, the fact that crispness levels were negatively associated with properties such as sound intensity or linear distance seems counterintuitive. In fact, several researchers have used such properties to assess levels of crispness. A key aspect here, however, is that panelists were cognizant that more than one term might be used to describe somewhat related textural attributes. Thus, if the evaluators were given only “crispness” as a descriptor, properties such as peak force or sound level might very well be correlated with sensory crispness levels. Machine learning models A predictive machine learning model was built based on sounds recorded during chip fracture. The prediction accuracy was tested using 20% of the data as the validation set (Table 6). The overall training performance of the model was 95.0%, that is the model accurately categorized 95% of the validation set. The training loss was 0.17 and measures the sum of errors for each example in the training set. A training performance greater than 70% is considered good model performance (Jiang 2021). In testing the model against the test set, the overall testing performance was 93.7%. Every chip category had an accuracy of greater than 89% and values were as high as 98% for the Kettle chips. Thus, the model was highly accurate in predicting the chip type from the sound produced while breaking. Table 6 Prediction of chip type based on sound by machine learning models Acc: 93.67% Lays Kettle Ruffle Dorito Noise Unknown Uncertain Lays 95% 0.50% 0% 1% 0% 0% 3.5% Kettle 0% 98% 1% 0% 0% 0% 1% Ruffle 2% 0% 90.5% 1.50% 0.5% 0% 5.5% Dorito 1.5% 1.5% 1.5% 89.5% 1.5% 0% 4.5% Noise 0.5% 0% 0% 0.5% 95% 1% 3% Unknown 0% 0% 0% 0% 5% 94% 1% F1 Score 0.95 0.98 0.94 0.93 0.94 0.96 While successfully identifying chips is useful, an additional goal of the research was to create a machine learning model that could predict the level of crispness and related attributes from the input of the sound of the snack food breaking. This was realized by using the sensory ratings from the descriptive panel as new labels for the created machine learning model. That is, the data labels such as “Lay’s Classic” or “Kettle” chips were changed to each chip’s corresponding trained panel rating. Thus, the model data were dubbed “Crispness 11.5”, “Crispness 10.5”, “Crispness 9.29” and “Crispness 8.14”. The overall testing performance of the “crispness” model was 92.6% (Table 7). Two additional models were created for the prediction of crunchiness and crackliness, in the same manner as the crispness model but using crunchiness and crackliness ratings for labeling. The overall performance of the crunchiness model was 94.5% (Table 8), while the overall performance of the crackliness model was 96.6% (Tables 9). Table 7 Prediction “crispness” ratings based on sound by machine learning models Acc: 92.58% Crisp 11.5 Crisp 10.5 Crisp 9.29 Crisp 8.14 Noise Unknown Uncertain Crisp 11.5 90.0% 0.0% 0.50% 2.0% 0.50% 0.0% 7.0% Crisp 10.5 0.50% 93.0% 2.50% 0.50% 0.0% 0.0% 3.50% Crisp 9.29 1.0% 0.0% 90.50% 3.0% 0.0% 0.0% 5.50% Crisp 8.14 3.0% 0.50% 0.50% 90.0% 1.50% 0.0% 1.0% Noise 0% 0% 0% 0.50% 97.0% 1.50% 1.0% Unknown 0% 0% 0% 0% 4.0% 95.0% 1.0% F1 Score 0.95 0.98 0.94 0.93 0.94 0.96 Table 8 Prediction of “crunchiness” ratings based on sound by machine learning models Acc: 94.5% Crunch 6.35 Crunch 11.6 Crunch 9.36 Crunch 9.78 Noise Unknown Uncertain Crunch 6.35 97.0% 1.0% 0.0% 1.0% 0.50% 0.0% 0.50% Crunch 11.6 0.50% 95.5% 1.50% 0.50% 0.0% 0.0% 2.0% Crunch 9.36 1.50% 0.0% 93.5% 3.50% 0.50% 0.0% 1.50% Crunch 9.78 1.0% 1.50% 1.0% 91.0% 2.0% 0.0% 3.50% Noise 1.0% 0.0% 0.0% 0.50% 97.0% 0.50% 1.50% Unknown 0.0% 0.0% 0.0% 0.0% 6.50% 93.0% 5.0% F1 Score 0.97 0.96 0.95 0.93 0.94 0.96 Table 9 Prediction of “crackliness” ratings based on sound by machine learning models Acc: 96.58% Crack 6.5 Crack 10.4 Crack 9.16 Crack 9.91 Noise Unknown Uncertain Crack 6.5 96.0% 0.0% 1.5% 0.0% 0.5% 0.0% 2.0% Crack 10.4 0.0% 99.0% 1.0% 0.0% 0.0% 0.0% 0.0% Crack 9.16 0.5% 0.0% 99.0% 0.0% 0.5% 0.0% 0.0% Crack 9.91 1.0% 0.0% 0.5% 92.5% 1.5% 0.0% 4.5% Noise 0.5% 0.0% 0.0% 0.50% 96.5% 1.0% 1.5% Unknown 0.0% 0.0% 0.0% 0.0% 3.0% 96.5% 0.5% F1 Score 0.99 0.97 0.98 0.96 0.96 0.98 There have been a few studies on the use of neural networks (NN) to assess food texture. Liu & Tan (1999) used specially designed pliers to crush several dry snacks, at two moisture levels, while recording sounds, and assessed crispness only on an unstructured scale from 0 to 10. They analyzed the sounds by signal value dependency, a frequency-based type of autocorrelation along with power value dependency , which plots how frequency spectra play out over time. PCA analysis showed 32 features survived screening although these were not explicitly stated. The neural network model was able to classify 10 samples into 4 crispness groups. Okada et al. (2016) used a tooth-like sensor that could detect load and vibrations, and developed an NN model that could classify snacks into several categories. Kato et al. (2019) simultaneously measured sound and load. The results were segmented into average loads in 5 time periods, and integration of frequency spectra in 5 frequency bands. Interestingly, they used convolution neural networks to analyze images of the spectra. They were able to estimate crispness and crunchiness on a scale of 0 to 1, but it is not clear how the scale was developed. Przybl et al. (2020) recorded the acoustic wave disturbance of dried strawberries rolling down a pipe. They found the optimal NN was a Multi-Layer Perceptron network that could classify firmness as measured by a texture analyzer. It might be pointed out that each approach for using sound to assess food texture has unique aspects in terms of how the measurement is made, what universe of foods are tested, what signal features are extracted for analysis and what type of neural network architecture is implemented. In our work we emphasized the simultaneous determination of crisp, crunchy and crackly attributes using a small, rounded probe that could fracture and continue through the sample. The sounds were further processed using MFCCs that have proven successful extracting features best suited for mimicking human hearing. In addition, this work used a trained-panel with a structured scale using specified standards for each texture attribute. Conclusion The goal of creating an ML model that could predict the type of chip based on the sound of chip breaking was realized. Furthermore, the model was extended to use sensorial crispness, crunchiness or crackliness ratings from trained panelists and was successfully able to identify how the chips were categorized as to levels of these attributes. However, at this point we cannot claim that the model could predict an intermediate level of crispness, crunchiness or crackliness. Mechanical and acoustical tests were performed on the chip snacks, also yielding data related to crispness, crunchiness and crackliness. Evaluations consistently showed the crispness levels were significantly distinct between each chip type. The data suggest that attributes such as peak force, linear distance or sound intensity might well predict levels of crunchiness or crackliness. Thus, it is still reasonable to conclude that these perceptions are related to the multiple fractures sensed in the mouth as well as the noisy sounds that are produced simultaneously. Crispness, however, was not readily predicted by combinations of physical attributes and this is likely due to panelists switching to crunchiness or crackliness as the relevant nomenclature as the value of these attributes increased. In this context, the use of machine learning models was valuable as they did very well at predicting the chip type or level of sensory attribute for all categories. Declarations Author contributions All authors contributed to the conception and design of the study. Toby Serrano: data collection, data curation, formal analyses, original draft. William Kerr- supervision and project administration, experimental design, data analysis, review and editing. Competing interest s The authors declare there are no financial or non-financial interests that are directly or indirectly related to the work submitted for publication. References Abdul ZK, Al-Talabani AK (2022) Mel frequency cepstral coefficient and its applications: a review. IEEE Access 10: 122144-122153. https://doi.org/10.1109/ACCESS.2022.3223444 Aleixandre A, Benavent-Gil Y, Velickova E, Rosell CM (2021) Mastication of crisp bread: role of bread texture and structure on texture perception. Food Res Int 147: 110477. https://doi.org/10.1016/j.foodres.2021.110477 Balemarthy S, Sajjanhar A, Zheng JX (2018) Our practice of using machine learning to recognize species by voice. arXiv Prepr. https://doi.org/10.48550/arXiv.1810.09078 Çarşanba E, Duerrschmid K, Schleining G (2018) Assessment of acoustic-mechanical measurements for crispness of wafer products. J Food Eng 229: 93-101. https://doi.org/10.1016/j.jfoodeng.2017.11.006 Castro-Prada EM, Primo-Martín C, Meinders MBJ, Hamer RJ, Van Vliet T (2009) Relationship between water activity, deformation speed, and crispness characterization. J Texture Stud 40: 152-171. https://doi.org/10.1111/j.1745-4603.2009.00173.x Chanvrier H, Jakubczyk E, Gondek E, Gumy JC (2014) Insights into the texture of extruded cereals: structure and acoustic properties. Innovative Food Sci Emerg Tech 24: 61-68. https://doi.org/10.1016/j.ifset.2013.11.013 Chaunier L, Courcoux P, Della Valle GUY, Lourdin D (2005) Physical and sensory evaluation of cornflakes crispness: a multidisciplinary approach. J Texture Stud 36: 93-118. https://doi.org/10.1111/j.1745-4603.2005.00007.x Chauvin MA, Younce F, Ross C, Swanson B (2008) Standard scales for crispness, crackliness and crunchiness in dry and wet foods: relationship with acoustical determinations. J Texture Stud 39: 345-368. https://doi.org/10.1111/j.1745-4603.2008.00147.x Cheely AN, Pegg RB, Kerr WL, Swanson RB, Huang G, Parrish DR, Kerrihard AL (2018) Modeling sensory and instrumental texture changes of dry-roasted almonds under different storage conditions. LWT-Food Sci Tech 91: 498-504. https://doi.org/10.1016/j.lwt.2018.01.069 Chen J, Karlsson C, Povey M (2005) Acoustic envelope detector for crispness assessment of biscuits. J Texture Stud 36: 139-156. https://doi.org/10.1111/j.1745-4603.2005.00008.x Dacremont C (1995) Spectral composition of eating sounds generated by crispy, crunchy and crackly foods. J Texture Stud 26: 27-43. https://doi.org/10.1111/j.1745-4603.1995.tb00782.x Das N, Mondal A, Chaki J, Padhy N, Dey N (2020) Machine learning models for bird species recognition based on vocalization: a succinct review. Inf Tech Intel Transp Sys 323: 117-124. https://doi.org/10.3233/FAIA200052 de Moraes JO, Andreani P, Murta BH, Link JV, Tribuzi G, Laurindo JB, Carciofi BA (2022) Mechanical-acoustical measurements to assess the crispness of dehydrated bananas at different water activities. LWT-Food Sci Tech 154: 112822. https://doi.org/10.1016/j.lwt.2021.112822 Hymel S (2022) Keyword spotting with edge impulse. GitHub, https://github.com/ShawnHymel/ei-keyword-spotting/tree/master. Accessed 10 February 2025 Jiang H (2021) Machine Learning Fundamentals: A Concise Introduction. Cambridge University Press Kato S, Wada N, Ito R, Shiozaki T, Nishiyama Y, Kagawa T (2019) Snack texture estimation system using a simple equipment and neural network model. Fut Internet 11: 68. https://doi.org/10.3390/fi11030068 Link JV, Tribuzi G, de Moraes JO, Laurindo JB (2018) Assessment of texture and storage conditions of mangoes slices dried by a conductive multi-flash process. J Food Eng 239: 8-14. https://doi.org/10.1016/j.jfoodeng.2018.06.024 Liu X, Tan J (1999) Acoustic wave analysis for food crispness evaluation. J Texture Stud 30: 397=408. https://doi.org/10.1111/j.1745-4603.1999.tb00227.x Lyons J (2009) Mel frequency cepstral coefficient (MFCC) tutorial. Practical cryptography, http://practicalcryptography.com/miscellaneous/machine-learning/guide-mel-frequency-cepstral-coefficients-mfccs/. Accessed 22 April 2025 Nimbarte N, Khan H, Sendre MD, Ramteke K, Wairagade S (2023) New born baby cry analysis and classification. 4 th Int Conf Emerging Tech (INCET). 2023: 1-5. https://doi.org/10.1109/INCET57972.2023.10170511 Okada S, Nakamoto H, Kobayashi F, Kojima F (2016) Intel Robotics App (ICIRA 2016). In: Kubota N, Kiguchi K, Liu H, Obo T (eds) Springer Cham, Switzerland, https://doi.org/10.1007/978-3-319-43506-0_21 Prabhakar H, Kerr WL, Bock CH, Kong F (2023) Effect of relative humidity, storage days, and packaging on pecan kernel texture: Analyses and modeling. J Texture Stud 54: 115-126. https://doi.org/10.1111/jtxs.12723 Primo-Martín C, De Beukelaer H, Hamer RJ, Van Vliet T (2008) Fracture behaviour of bread crust: Effect of bread cooling conditions. J Food Eng 89: 285-290. https://doi.org/10.1016/j.jfoodeng.2008.05.005 Przybył K, Duda A, Koszela K, Stangierski J, Polarczyk M, Gierz L (2020) Classification of dried strawberry by the analysis of the acoustic sound with artificial neural networks. Sensors 20: 499. https://doi.org/10.3390/s20020499 Rodriguez CR, Angeles D, Chafloque R, Kaseng F, Pandey B (2020) Deep learning audio spectrograms processing to the early COVID-19 detection. 12th Int Conf Comp Intel Comm Netw. 2020: 429-434. https://doi.org/10.1109/CICN49253.2020.9242583 Roudaut G, Dacremont C, Meste ML (1998) Influence of water on the crispness of cereal-based foods: acoustic, mechanical, and sensory studies. J Texture Stud 29: 199-213. https://doi.org/10.1111/j.1745-4603.1998.tb00164.x Roudaut G, Dacremont C, Pàmies BV, Colas B, Le Meste M (2002) Crispness: a critical review on sensory and material science approaches. Trends Food Sci 13: 217-227. https://doi.org/10.1016/S0924-2244(02)00139-5 Salvador A, Varela P, Sanz T, Fiszman SM (2009) Understanding potato chips crispy texture by simultaneous fracture and acoustic measurements, and sensory analysis. LWT-Food Sci Tech 42: 763-767. https://doi.org/10.1016/j.lwt.2008.09.016 Stone H, Bleibaum RN, Thomas HA (2020) Sensory Evaluation Practices. Academic Press, New York Suwonsichon T, Peleg M (1998) Instrumental and sensory detection of simultaneous brittleness loss and moisture toughening in three puffed cereals. J Texture Stud 29: 255-274. https://doi.org/10.1111/j.1745-4603.1998.tb00169.x Szczesniak AS, Kahn EL (1971) Consumer awareness of and attitudes to food texture: I. Adults. J Texture Stud 2: 280-295. https://doi.org/10.1111/j.1745-4603.1971.tb01005.x Szczesniak AS, Kahn EL (1984) Texture contrasts and combinations: a valued consumer attribute. J Texture Stud 15: 285-301. https://doi.org/10.1111/j.1745-4603.1984.tb00385.x Taniwaki M, Kohyama K (2012) Mechanical and acoustic evaluation of potato chip crispness using a versatile texture analyzer. J Food Eng 112: 268-273. https://doi.org/10.1016/j.jfoodeng.2012.05.015 Trivedi K, Shroff H (2021) identification of deadliest mosquitoes using wing beats sound classification on tiny embedded system using machine learning and Edge Impulse platform. ITU Kaleidosc Conn Phys Virt Worlds. 2021: 1-6. https://doi.org/10.23919/ITUK53220.2021.9662116 Tunick MH, Onwulata CI, Thomas AE, Phillips JG, Mukhopadhyay S, Sheen S, Cooke PH (2013) Critical evaluation of crispy and crunchy textures: a review. Int J Food Props 16: 949-963. https://doi.org/10.1080/10942912.2011.573116 Varela P, Chen J, Fiszman S, Povey MJW (2006) Crispness assessment of roasted almonds by an integrated approach to texture description: texture, acoustics, sensory and structure. J Chemometrics 20: 311-320. https://doi.org/10.1002/cem.1029 Vicker Z M (1984a) Crackliness: relationships of auditory judgments to tactile judgments and instrumental acoustical measurements. J Food Tex 15: 49-58. Vickers ZM (1984b) Crispness and crunchiness—a difference in pitch? J Texture Stud 15: 157-163. https://doi.org/10.1111/j.1745-4603.1984.tb00375.x Vincent JF (1998) The quantification of crispness. J Sci Food Agr 78: 162-168. https://doi.org/10.1002/(SICI)1097-0010(199810)78:23.0.CO;2-3 Xu S, Kerr WL (2012) Comparative study of physical and sensory properties of corn chips made by continuous vacuum drying and deep fat frying. LWT-Food Sci Tech 48: 96=101. https://doi.org/10.1016/j.lwt.2012.02.019 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6514631","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":449994872,"identity":"9907b9c1-930b-4c5b-9f0d-6a1b6303c5ef","order_by":0,"name":"Toby Serrano","email":"","orcid":"","institution":"University of Georgia","correspondingAuthor":false,"prefix":"","firstName":"Toby","middleName":"","lastName":"Serrano","suffix":""},{"id":449994873,"identity":"3661d498-f0dc-400a-8c41-42519ef60a32","order_by":1,"name":"William Kerr","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAArklEQVRIiWNgGAWjYDACdjYQaQPh8BClhRmsJY10LYdJ0GJwmC3xceWe84nz3Q8wPnjbRpyWw4Znnt1O3HgmgdlwLjFazA6zt0k2HABqaUhgk+YlUkv7z4YD5xI39j9g/02kFrZjjA0HDiTOl0hgYyZKi/1htmSgw5KNN0g8bJacc44ILZLtbYYfGw7Yyc7vTz744U0ZEVrgwOAAYwMp6oFAnlQNo2AUjIJRMHIAAC5fONK4QxcSAAAAAElFTkSuQmCC","orcid":"","institution":"University of Georgia","correspondingAuthor":true,"prefix":"","firstName":"William","middleName":"","lastName":"Kerr","suffix":""}],"badges":[],"createdAt":"2025-04-23 17:08:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6514631/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6514631/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":81950864,"identity":"72b34c7a-1e92-4476-be3f-6eed06fc8ad4","added_by":"auto","created_at":"2025-05-05 09:12:22","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":150502,"visible":true,"origin":"","legend":"\u003cp\u003eQuantitative sensory descriptive ratings for Lay’s\u003csup\u003e®\u003c/sup\u003e classic potato chips (¾), Lay’s\u003csup\u003e®\u003c/sup\u003e kettle cooked potato chips (----); Ruffles\u003csup\u003e® \u003c/sup\u003eoriginal potato chips (\u003cstrong\u003e¾\u003c/strong\u003e); and Doritos\u003csup\u003e®\u003c/sup\u003e nacho cheese tortilla chips (- - -)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6514631/v1/3b6124f5a18ee03a7dbcc1fd.png"},{"id":81952508,"identity":"7783e96c-bada-44a3-850b-97f60c1baabc","added_by":"auto","created_at":"2025-05-05 09:28:22","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":278537,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6514631/v1/bfa08419046be394f221473c.png"},{"id":92161812,"identity":"f59c3e37-1383-4d5a-aa25-1a4d34370019","added_by":"auto","created_at":"2025-09-25 10:09:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1044152,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6514631/v1/514f69a6-2759-4d47-8768-c014a19a1692.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Machine Learning Models of Sounds to Predict Textural Attributes of Snack Chips","fulltext":[{"header":"Introduction","content":"\u003cp\u003eTexture is an important attribute of many food products, directly influencing consumer preferences and marketability (Szczesniak\u0026nbsp;and Kahn 1971).\u0026nbsp;However, there are specific texture properties that present as foods break abruptly into several pieces when chewed and are accompanied by complex sounds. Several terms are used to describe these attributes, and they may be used interchangeably by consumers.\u0026nbsp;\u0026ldquo;Crispness\u0026rdquo;\u0026nbsp;is one of these and regarded as one of the most desired attributes in food. Szczesniak and Kahn (1984) found that consumers view\u0026nbsp;crispness\u0026nbsp;as well-liked, a sign of excellent cooking, and \u0026ldquo;very synonymous with freshness and wholesomeness\u0026rdquo;. They also showed that\u0026nbsp;crispness\u0026nbsp;was versatile, in that it could be paired with other texture attributes such as softness, smoothness, and creaminess.\u003c/p\u003e\n\u003cp\u003e\u0026ldquo;Crunchiness\u0026rdquo;\u0026nbsp;is a related textural attribute enjoyed nearly as much as\u0026nbsp;crispness. In their research Szczesniak \u0026amp; Kahn (1984) indicated that\u0026nbsp;crunchiness\u0026nbsp;adds contrast and liveliness to food texture. As with\u0026nbsp;crispness,\u0026nbsp;crunchiness\u0026nbsp;is desirable in combination with many other textures. The public perception of\u0026nbsp;crunchiness\u0026nbsp;is highly positive and \u0026ldquo;regarded with warmth and affection\u0026rdquo;. \u0026nbsp;\u0026ldquo;Crackliness\u0026rdquo;\u0026nbsp;is another textural attribute found in snack foods but has not been as well studied as\u0026nbsp;crispness\u0026nbsp;or\u0026nbsp;crunchiness.\u0026nbsp;Crackliness\u0026nbsp;can be defined as \u0026ldquo;multiple sharp repeated noises when a food is bitten and chewed\u0026rdquo; (Vickers 1984a).\u003c/p\u003e\n\u003cp\u003eManufacturers and researchers have struggled to find methods to determine the quality and intensity of crisp, crunchy or crackly texture attributes. Of course, the primary means of assessment is through sensory analyses performed by human subjects. This may involve untrained panelists that are tasked to determine the level of likability or preference of one product over another. In contrast, trained or experienced persons can be used to determine the intensity level of one or more attributes such as crunchiness, hardness or chewability.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInstrumental texture analysis is valuable for quickly assessing changes in quality, comparing products or determining the effects of new formulations (Roudaut et al. 2002). This typically involves compressive, shear or puncture probes that are lowered through the sample, and the force-distance response measured as the sample deforms and breaks.\u0026nbsp;The crispness of many different types of foods has been analyzed using such mechanical methods. This includes bread crust (Primo-Mart\u0026iacute;n et al. 2008), potato chips (Salvador et al. 2009), wafers (\u0026Ccedil;arşanba et al. 2018), crisp bread (Aleixandre et al. 2021), almonds (Cheely et al. 2018) and extruded cereals (Chanvrier et al. 2014), as well as dried fruits such as mango slices (Link et al. 2018) and bananas (de Moraes et al. 2022).\u003c/p\u003e\n\u003cp\u003eOne challenge is how to best analyze rather noisy data to provide a metric that reflects crispness and similar attributes. For simple systems such as saltine crackers or almonds (Cheely et al. 2018), this may be a single maximum or \u0026ldquo;snap\u0026rdquo; force that correlates with crispness. Likewise, the force or distance to the first fracture event may be deemed important. Many approaches take note of the numerous fracture events that occur under stress. Thus, Prabhakar et al. (2023) found the ratio of the initial fracture force to peak force useful in assessing pecan crunchiness. Others have enumerated the total number of peaks during breakage. Suwonsichon and Peleg (1998) studied fluctuations of the force-displacement relationship, as quantified by its fractal dimension, which was observed to change with varying moisture contents in the cereals. Another approach is to measure the deviations by summing the total distances for each point-to-point length, as was done for corn chips (Xu and Kerr 2012). If force data is collected over time such that there are \u003cem\u003en\u003c/em\u003e data points, the \u0026ldquo;linear distance\u0026rdquo; is given by:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003eCrisp, crunchy and crackly foods generally make sounds as they are consumed. Dacremont (1995) recorded both air and bone conducted sounds as panelists chewed a variety of crisp, crunchy or crackly foods. Chen et al. (2005) described an \u0026ldquo;Acoustic Envelope Detector\u0026rdquo; attached to a texture analyzer for concurrently obtaining acoustic and force data from food samples. For biscuits, it was found that for every force drop, there was a simultaneous auditory occurrence. When studying the effects of water activity Castro-Prada et al. (2009) found that greater sound intensity related to a higher level of\u0026nbsp;\u0026ldquo;crispness\u0026rdquo;.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMultiple acoustic parameters have been calculated to assess textural properties. In research on wafers (\u0026Ccedil;arşanba et al. 2018), potato chips (Salvador et al. 2009;\u0026nbsp;Taniwaki and Kohyama 2012) and almonds (Varela et al. 2006), greater maximum sound pressure (S\u003csub\u003emax\u003c/sub\u003e) and number of acoustic peaks in the audio waveform were positively correlated with sensory\u0026nbsp;crispness. When studied simultaneously, it has been noted that the fluctuations in both mechanical and acoustical curves are an indication of\u0026nbsp;crispness\u0026nbsp;(Salvador et al. 2009; Link et al. 2018; Chen et al. 2005; Vincent 1998).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eResearch has also included frequency analysis, particularly when comparing crispy to crunchy foods. Dacremont (1995) studied crispy food samples including flatbreads, while the crunchy samples included foods such as raw carrots. They found that crisp foods generate more high frequency sounds (\u0026gt;5kHz), crunchy foods have more sound energy at \u0026lt;1.7 kHz and crackly foods generate low pitched sounds with a high level of bone conduction.\u003c/p\u003e\n\u003cp\u003eA promising approach to categorizing sounds is through machine learning (ML). This consists of recording relevant audio events, processing these to generate useful features while reducing data size (Lyons 2009), then using these to train models using large datasets. Training consists of analyzing the data with neural network algorithms, recognizing the data patterns, and using the resulting set of nested equations to identify unknown items. For example, Nimbarte et al. (2023) used Python to build an ML model to detect crying babies, with an accuracy of 0.8 and precision of 0.48. An ML model was developed with Google Colab to detect COVID-19 cases from the audio spectrograms of coughing, sneezing, and respiratory sounds (Rodriguez et al. 2020). ML has also been used to\u0026nbsp;recognize various animal species (Balemarthy et al. 2018) and birds (Das et al. 2020) based on their vocalizations. Trivedi and Shroff (2021) utilized an Arduino microcomputer and Edge Impulse to classify the species of deadly mosquitoes based on recorded audio data of mosquito wing beats. For food research, Kato et al. (2019) developed an ML-based system to estimate crispness from force and sound data. However, there was no inputs of human perceptions of crispness.\u003c/p\u003e\n\u003cp\u003eIn this research, an ML model was trained to classify the sounds of different snack chips, with the goal that after training, the model should be able to determine the kind of snack food based on the sound it emits during compression. Concurrently, the snack foods were evaluated for their textural attributes using sensory assessments, force-time response, and traditional acoustic calculations, to provide greater understanding of the causes of any differences. In addition, ML models were constructed to differentiate and predict levels of crispness, crunchiness or crackliness based on sensory evaluations of each of these attributes.\u0026nbsp;\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003e\u003cstrong\u003eSamples\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe chips used were Lay's\u003csup\u003e®\u0026nbsp;\u003c/sup\u003eClassic Potato Chips, Lay's\u003csup\u003e®\u0026nbsp;\u003c/sup\u003e Kettle Cooked Original Potato Chips, Ruffles\u003csup\u003e®\u0026nbsp;\u003c/sup\u003e Original Potato Chips, and Doritos\u003csup\u003e®\u003c/sup\u003e Nacho Cheese Flavored Tortilla Chips (Frito-Lay, Plano, Texas), purchased from a local grocery store in Athens, Georgia. The chip samples had minimal preparation before analysis and were analyzed immediately after removal from the original packaging to retain freshness.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eSensory analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll samples were evaluated by trained panel using quantitative descriptive analysis (Chauvin et al. 2008; Stone et al. 2020). Nine participants (3M, 6F; age range 20 to 60 y) were recruited from the University of Georgia and operated under protocols approved by the UGA Institutional Review Board (PROJECT00006724). Panelists received a small compensation for their participation in the study.\u003c/p\u003e\n\u003cp\u003eThe sensory analysis was conducted through six different sessions, each held on a different day. In the first session, panelists were screened for hearing, dental, or other issues that would prevent consumption of hard snack foods. Panelists were taught to sample in a specific manner as follows: hold one piece of the sample, take one bite of the sample using the incisor teeth, and then chew using the molar teeth and swallow; after each bite, rinse the mouth with filtered water, then expectorate the rinse water into a provided container. During training, participants discussed textural attributes that best described all the samples and concluded that the attributes “crispness”, “crunchiness”, and “crackliness” were the essential textural attributes of the chip snacks. Panelists were then provided definitions of “crispness”, “crunchiness”, and “crackliness” as described in Chauvin et al. (2008). The group next developed a new standard scale for each of the textural attributes. The foods were prepared by placing each sample in a 2-oz lidded plastic cup.\u003c/p\u003e\n\u003cp\u003eIn the next session, panelists were asked to rate each food standard on a 15-point scale, with 15 being the most extreme level of the textural attribute, and 0 being a complete lack of that attribute. Panelists rated each food first by themselves, writing the name of each food on a point of the line. Panelists afterward discussed among themselves to agree in unanimity on where each food belongs on the 15-point line.\u003c/p\u003e\n\u003cp\u003eIn subsequent sessions, panelists were asked to rate the four different chip samples on the three standard scales for crispness, crunchiness, and crackliness. At this stage, evaluations occurred in isolated sensory booths. Each of the different chip types were labeled with a randomized three-digit code. Panelists were provided with a sample of each food, a napkin, rinsing water, an expectorant container, a pen, and a paper with the three developed standard scales. Standards were also provided so panelists could remind themselves of the standard scale. In the fourth, fifth, and sixth sessions, panelists again rated the four different kinds of chips. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInstrumental texture analyses\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe four chip samples were analyzed using a TA.XT Plus texture analyzer (Stable Micro Systems, Surrey, UK) fitted with a “Crisp Fracture Support Rig”. Chip samples were placed on a cylindrical stand 25 mm in diameter as a 6.2 mm ball tip probe descended through the center of the sample at 5 mm/s. The force versus time data was analyzed using the software, determining the peak force in Newtons (N), the linear distance (AU), the area under the curve in Ns, and the number of peaks in the curve using a threshold of 68.6 mN. At least 200 replications of chps of each type were analyzed. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcoustic analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn Arduino Nano 33 BLE Sense microcomputer connected to a laptop was positioned 10 mm away from the chip sample. The Arduino’s PDM microphone recorded the sounds during the crushing and puncture of the chip sample by the probe. At least 200 unique audio files were recorded for each of the chip types. Using an in-house MATLAB (R2023b) script, the audio files were analyzed for their sound intensity, number of sound peaks, and percentage of cumulative frequencies under 1.7 kHz (Mathwork Inc, Nattick, MA).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;Machine learning models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe “.wav” files were trimmed to 5 s each and each was given a label for its corresponding type of chip. At least 200 unique audio files were recorded for each chip type. To prevent overfitting, the sets of audio files were processed in Python 3.10 using the Google Speech Commands dataset (Hymel 2022). This data augmentation tool mixed the chip audio files with background noise to create more audio files for each chip type. With this procedure, 800 augmented chip audio files were created, totaling 1000 chip audio files for each chip type: 4000 in total. In addition, audio samples of sound representing “unknown” sound and “noise” were obtained from Shawn Hymel’s GitHub programming code repository, part of Edge Impulse’s Coursera online learning platform (Hymel 2022). For both “unknown” and “noise”, 1000 audio files were obtained: 2000 in total.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe 6000 audio files comprising direct chip sounds, ‘noise’ and ‘unknown’ sounds (totaling 8 h and 20 min) of audio, were imported into the Edge Impulse platform. The entire dataset was randomly split into one training and one test set. The training set encompassed 80 percent of the data (4800 files), while the test set encompassed 20 percent of the data (1200 files). The files were then edited in the processing block. The ‘window size’, that is the amount of data to be converted for each classification, was set to the length of each audio file, that is 5000 ms. The ‘window increase’ (used to artificially create more features) was 500 ms, the frequency of the data was 16000 Hz, and zero-pad setting was enabled.\u003c/p\u003e\n\u003cp\u003eThe Mel frequency cepstral coefficients technique was used for feature extraction. The MFCCs represents a reduced dataset that captures critical temporal and frequency information of the audio signal that is most relevant to the model. Research has shown that MFCCs are effective in classifying sounds as they consider sound energy in frequency bands that mimic the human auditory response (Abdul and Al-Talabani 2022). The parameters used for the MFCC design were as follows. The number of cepstral coefficients was 13; the length of each frame was 20 ms; the frame stride was 20 ms; the number of filters in the filter bank was 32; the FFT length was 256; the normalization window size was 101; and the lowest frequency was 0 Hz. The high frequency was set to half of the sample rate and the pre-emphasizing coefficient was 0.98.\u003c/p\u003e\n\u003cp\u003eFor neural network training, the number of training cycles was 100 and the learning rate set to 0.005. The validation set was 20% of the training set. The Profile int8 model setting was turned on. This is a quantization tool that converts data to 8-bit integers instead of floating-point numbers and integer math instead of floating-point math, thus reducing memory and computing requirements. The first layer was an Input layer of 3250 features. The second layer was a Reshape layer made up of 13 columns. The third layer was a 1-dimensional single convolution layer followed by a pooling layer, made of 8 neurons and a kernel size of 3. The fourth layer was a Dropout layer, with a dropout rate of 0.25. The fifth layer was a 1-dimensional single convolution/pool layer of 25 neurons and a kernel size of 25. The sixth layer was another Dropout layer with a 0.25 dropout rate. The seventh layer was a Flatten layer. The eighth and final layer was the output layer, with the 6 different classes of Lays Classic, Kettle, Ruffles, Doritos, “unknown”, and “noise”.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical analyses\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;The data were initially analyzed using the Shapiro-Wilks test (SPSS 29 Statistics software, Armank, NY). The test is based on the null hypothesis that the data set is normally distributed and outputs a value from 0 to 1, where 1 indicates the data closely resembles the normal distribution. Thus, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.05 signifies that the assumption of normality must be rejected. In concert, data were tested for homogeneity of variance using Levene’s test. Here, p\u0026lt;0.05 indicates that the null hypothesis of equal variances is rejected and there is a difference between the variances in the population. When normality and equal variances could not be assured, the Kruskal-Wallis was used to test whether samples originated from the same distribution. Thus, it serves as the non-parametric equivalent of the one-way ANOVA. When differences were found, the post-hoc Dunn’s test was used as a pairwise comparison test that assesses the differences between group medians. Principal component analysis was done using JMP 12 (JMP Statistical Discovery LLC, Cary, NC).\u003c/p\u003e"},{"header":"Results and discussion","content":"\u003cp\u003e\u003cstrong\u003eSensory analysis\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eResults from the modified quantitative descriptive analyses of the chips are shown in Figure 1. The highest crispness intensity (11.5) was found for the Lay’s Classic potato chips and lowest for the Doritos tortilla chips (8.14). The greatest crunchiness (11.6) and crackliness (10.4) was noted for the Lay’s Kettle Cooked potato chips. For Lay’s Classic chips the crispness (at 11.5) was clearly the predominant characteristic with scores of 6.35 and 6.5 for crunchiness and crackliness, respectively. Some researchers have suggested that foods perceived as more crisp than crunchy are less dense (Tunick et al. 2013), and this may be why the Lay’s Classic chips were perceived as more crisp than crunchy. For the other chips, the values of the texture attributes were more similar. Thus, for the Lay’s Kettle Cooked chips values of crispness, crunchiness and crackliness were 10.5, 11.6 and 10.4, while for the Ruffles, these values were 9.29, 9.36 and 9.16. This shows that these textural terms are not disconnected, and even trained panelists can perceive more than one attribute for each food item. However, for some foods one of the textural percepts may predominate. With this understanding, some researchers have even used sensory ballots with terminology such as “more crisp than crunchy” or “more crunchy than crisp” (b 1984). \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe Shapiro-Wilk test showed p\u0026lt;0.05 for all crispness evaluations, as well as for crunchiness and crackliness for the Lay’s Classic and Kettle Cooked. Thus, most data could not be said to be normally distributed.\u0026nbsp;Further, Levene’s test showed that only crispness ratings had a \u003cem\u003ep\u003c/em\u003e\u0026gt; 0.05 so that only the crispness ratings had homogeneity of variance. This may be traced to the fact that the chips did not have uniform shape and size. Because of the lack of normality and homogeneity of variance, non-parametric tests for significant differences were used.\u003c/p\u003e\n\u003cp\u003eThe Kruskal-Wallis test showed that\u003cem\u003e\u0026nbsp;p\u003c/em\u003e\u0026lt;0.05 for crispness, crunchiness, and crackliness ratings indicating the likelihood of differences amongst the attributes. The Dunn’s test was subsequently used to test for significant differences between paired groups (Table 1). For crispness, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.05 when comparing each chip against another. Thus, each chip had a distinguishable level of crispness compared to the others. For crunchiness, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.05 for all comparisons except Kettle-Dorito and Ruffle-Dorito, indicating that Kettle or Ruffles did not have different crunchiness than Doritos. For crackliness, \u003cem\u003ep\u003c/em\u003e\u0026gt;0.05 for Kettle-Ruffle, Kettle-Dorito, and Ruffle-Dorito, indicating that crackliness ratings were not very distinct amongst chips.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e Dunn’s test for significant difference in sensory descriptive ratings between\u003c/p\u003e\n\u003cp\u003esnack chip types\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"525\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrispness\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunchiness\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrackliness\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eChip Type\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eLays-Kettle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eLays-Ruffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eLays-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eKettle-Ruffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.026\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.012\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.567\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eKettle-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.092\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eRuffle-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.031\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe results suggest that the panelists’ understanding of crispness was the strongest, leading to a clear difference in ratings between chip types, while the understanding of crunchiness and crackliness was vaguer. Another possibility of course is that the perceived crunchiness and crackliness between Kettle Cooked, Ruffles and Doritos Chips are very similar, leading to an unclear difference in ratings.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInstrumental texture analyses\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eResults from the instrumental texture analyses are shown in Table 2.\u0026nbsp;The\u0026nbsp;Kettle Cooked chips showed the greatest peak force (8.32 N), followed by the Doritos (6.49 N), Ruffles (5.07 N) and Lay’s Classic chips (2.75 N). The Kettle Cooked also had the greatest number of fracture peaks (28.4), followed by Doritos (16.8), Ruffles (16.2) and Lay’s Classic (13.2). Finally, the linear distance decreased in the order: Kettle Cooked (65.5 Ns), Doritos (46.3 Ns), Ruffles (40.2 Ns) and Lay’s Classic (30.1 Ns). Thus, the order of decrease was the same for each force-time metric derived from the data. As noted, the Lay’s Classic were perceived as having the greatest sensory rating of crispness yet were lowest in terms of force-related attributes. These force-related metrics have been correlated with crispness in other research (Aleixandre et al. 2021; Suwonsichon\u0026nbsp; and Peleg, 1998; Chaunier et al. 2005; Roudaut et al. 1998). In this case, it seems that texture analyzer values would not be the best predictor of crispness when crunchiness or crackliness are considered at the same time.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e Analyses of force-time response during fracturing of snack chips\u003c/p\u003e\n\u003cp\u003eby instrumental texture analyzer.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"529\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eChip Type\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePeak Force (N)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLinear Distance (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eArea Under Curve (Ns)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNumber of Peaks\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLays\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e30.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e13.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStd Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(0.87)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(5.72)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(2.07)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(7.93)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKettle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e8.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e65.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e10.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e28.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStd Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(3.94)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(33.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(9.03)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(16.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRuffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e40.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e16.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStd Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(1.74)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(10.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(3.72)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(8.85)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6.49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e46.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e16.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStd Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(1.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(13.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(3.34)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e(7.96)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe Shapiro-Wilk and Levene’s tests all had \u003cem\u003ep\u003c/em\u003e\u0026lt;0.05, again indicating the data were not normally distributed nor had homogeneity of variance. The Kruskal-Wallis \u003cem\u003ep\u003c/em\u003e-values \u0026lt; 0.05 for every parameter, indicating there were differences in mechanical properties amongst the chip types. The Dunn’s test (Table 3) showed\u0026nbsp;\u003cem\u003ep\u003c/em\u003e\u0026lt;\u0026nbsp;0.05, except for the peak force of “Kettle-Dorito” (p=0.116) and number of peaks for “Ruffle-Dorito” (p=0.99). Thus, except for a few comparisons, the force parameter values were significantly different between every chip type.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eTable 3\u003c/strong\u003e Dunn’s test for significant difference in force parameters between specific chips\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"572\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePeak Force\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLinear Distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eArea Under Curve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNumber of Peaks\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eChip Type\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLays-Kettle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLays-Ruffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLays-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKettle-Ruffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKettle-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.116\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRuffle-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eAcoustical analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eResults from the acoustic analyses of sounds recorded during chip breakage are shown in Table 4. In order, the overall sound intensity decreased from\u0026nbsp;Kettle Cooked\u0026nbsp;(9770 au),\u0026nbsp;Doritos (8616 au), Ruffles (8070 au) and\u0026nbsp;Lay’s Classic\u0026nbsp;(7776 au). The number of sound peaks followed the same order ranging from 34.8 for the\u0026nbsp;Kettle Cooked\u0026nbsp;chips to 16.6 for the\u0026nbsp;Lay’s Classic. Interestingly, this also is the same order found for the peak force and number of force peaks. This reinforces the idea that the sound peaks emanate from individual fracture advents while the sound level is tied to the energy released during the mechanical fracture.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4\u003c/strong\u003e Analyses of sound recorded during fracturing of snack chips\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"446\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSound\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eIntensity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNumber of Sound Peaks\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e% \u0026lt; 1.7 kHz\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKettle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e9,770\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e34.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e30.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStd Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2,012\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e7.34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e8,616\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e31.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStd Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1,663\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRuffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e8,070\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e17.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e22.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStd Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e968\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e7.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLays\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e7,777\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e16.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e23.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStd Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e929\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe sound recordings were also analyzed by Fourier transform algorithms to determine the power spectrum density, that is the amount of acoustical energy associated with each frequency. Thus, the analyses plot the relative weights of frequency components that make up the audio signal. Some work has suggested that items sensed as more crunchy than crispy have more acoustical energy at lower frequencies, that is the signal contain more lower “pitches” (Dacremont 1995; Vickers 1984). The percentage of the cumulative frequency under 1.7 kHz for each chip type was also calculated (Table 4). This marker was chosen as snack items with frequencies higher than 1.7 kHz were associated with\u0026nbsp;“crispness”, while frequencies lower than 1.7 kHz were associated with\u0026nbsp;“crunchiness”\u0026nbsp;(Dacremont 1995).\u0026nbsp;Kettle Cooked\u0026nbsp;and Dorito chips both had a percent under 1.7 kHz around 31%, while\u0026nbsp;Lay’s Classic\u0026nbsp;and\u0026nbsp;Ruffles Original\u0026nbsp;had a percentage of 23.2 and 22.3, respectively. This at least concurs with the observation that\u0026nbsp;Lay’s Classic\u0026nbsp;chips were perceived as the crispest while\u0026nbsp;Kettle Cooked\u0026nbsp;were seen as most crunchy, as more of the sound energy resides at higher frequencies.\u003c/p\u003e\n\u003cp\u003eThe Shapiro-Wilk test showed \u003cem\u003ep\u003c/em\u003e\u0026gt;0.05 for every chip when considering sound intensity, indicating a more normal distribution for this data set. For number of sound peaks \u003cem\u003ep\u003c/em\u003e\u0026lt; 0.05, except for Doritos, thus suggesting a non-normal distribution of values. The Levene’s test showed \u003cem\u003ep\u003c/em\u003e\u0026lt; 0.05 for sound intensity, but \u003cem\u003ep\u003c/em\u003e\u0026gt;0.05 for number of peaks, showing that while the sound intensity values might be considered normally distributed, they did not have homogeneity of variance. The Kruskal-Wallis tests indicated there were significant differences between chip types for the sound properties. Dunn’s Test (Table 5) showed that for sound intensity \u003cem\u003ep\u003c/em\u003e\u0026lt; 0.05 between all chip comparisons. Furthermore, the number of sound peaks were different between all types except between “Ruffle-Dorito”. Thus except for one comparison, each chip was distinct in their sound intensity and number of sound peaks.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 5\u003c/strong\u003e Dunn’s test for significant difference in sound parameters between chip types\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"478\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSound Intensity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNumber of Sound Peaks\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eChip Types\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLays-Kettle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLays-Ruffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLays-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKettle-Ruffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKettle-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt;0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRuffle-Dorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.135\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAs shown in Figure 1, Table 2 and Table 4,\u0026nbsp;crispness\u0026nbsp;ratings did not follow the same pattern as the mechanical and acoustic parameters. That is, the\u0026nbsp;crispness\u0026nbsp;ratings, from highest to lowest, were\u0026nbsp;Lay’s Classic,\u0026nbsp;Kettle Cooked,\u0026nbsp;Ruffles, and Doritos. The mechanical and acoustic parameters were ranked\u0026nbsp;Kettle Cooked, Doritos, Ruffles and Lays Original chips. However, the\u0026nbsp;crunchiness\u0026nbsp;ratings had the exact same ordering as the mechanical and acoustic parameters.\u003c/p\u003e\n\u003cp\u003eThis was further analyzed through PCA plots (Figure 2). Thus, crunchiness and to a lesser extent crackliness, were most associated with high peak force and linear distance measurements. The linear correlations for crunchiness and peak force were r\u003csup\u003e2\u003c/sup\u003e=0.98 and for linear distance r\u003csup\u003e2\u003c/sup\u003e=0.94. For crackliness, these were r\u003csup\u003e2\u003c/sup\u003e=0.95 and 0.85, respectively. Other metrics were reasonable predictors of crunchiness and crackliness. Thus, crunchiness was linearly related to the number of force peaks (r\u003csup\u003e2\u003c/sup\u003e=0.86) as well as sound intensity (r\u003csup\u003e2\u003c/sup\u003e=0.79). Crunchiness was most associated with greater acoustic energy at low frequencies, \u0026nbsp;measured by %\u0026lt;1.7kHz. It should be noted that many of the force and acoustic measurements are interdependent, thus each might reasonably serve as a proxy for texture measurement. For example, the number of force peaks was correlated with the number of sound peaks (r\u003csup\u003e2\u003c/sup\u003e=0.99) and overall sound intensity (r\u003csup\u003e2\u003c/sup\u003e=0.97). Likewise, the linear force distance was related to number of peaks or sound intensity (r\u003csup\u003e2\u003c/sup\u003e=0.98)\u003c/p\u003e\n\u003cp\u003eAn interesting observation is that crispness was negatively correlated with most other physical measurements. The greatest negative correlation was with crackliness (r\u003csup\u003e2\u003c/sup\u003e=-0.64) and crunchiness (r\u003csup\u003e2\u003c/sup\u003e=-0.44). In terms of physical measurements, crispness was most negatively correlated to peak force (r\u003csup\u003e2\u003c/sup\u003e=-0.40) and %\u0026lt;1.7kHz (r\u003csup\u003e2\u003c/sup\u003e=-0.38). The theory that crisp foods are more associated with higher frequency content and crunchy/crackly foods have more low frequencies is borne out by the results. However, the fact that crispness levels were negatively associated with properties such as sound intensity or linear distance seems counterintuitive. In fact, several researchers have used such properties to assess levels of crispness. A key aspect here, however, is that panelists were cognizant that more than one term might be used to describe somewhat related textural attributes. Thus, if the evaluators were given only “crispness” as a descriptor, properties such as peak force or sound level might very well be correlated with sensory crispness levels.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMachine learning models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA predictive machine learning model was built based on sounds recorded during chip fracture. The prediction accuracy was tested using 20% of the data as the validation set (Table 6). The overall training performance of the model was 95.0%, that is the model accurately categorized 95% of the validation set. The training loss was 0.17 and measures the sum of errors for each example in the training set. \u0026nbsp;A training performance greater than 70% is considered good model performance (Jiang 2021). In testing the model against the test set, the overall testing performance was 93.7%. Every chip category had an accuracy of greater than 89% and values were as high as 98% for the Kettle chips. Thus, the model was highly accurate in predicting the chip type from the sound produced while breaking.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 6\u003c/strong\u003e Prediction of chip type based on sound by machine learning models\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"576\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eAcc:\u003c/p\u003e\n \u003cp\u003e93.67%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eLays\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eKettle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eRuffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eDorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eNoise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eUncertain\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eLays\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e95%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e3.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eKettle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e98%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eRuffle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e90.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e5.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eDorito\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e89.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e4.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eNoise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e95%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e94%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e1%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003eF1 Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eWhile successfully identifying chips is useful, an additional goal of the research was to create a machine learning model that could predict the level of crispness and related attributes from the input of the sound of the snack food breaking. This was realized by using the sensory ratings from the descriptive panel as new labels for the created machine learning model. That is, the data labels such as “Lay’s Classic” or “Kettle” chips were changed to each chip’s corresponding trained panel rating. Thus, the model data were dubbed “Crispness 11.5”, “Crispness 10.5”, “Crispness 9.29” and “Crispness 8.14”. The overall testing performance of the “crispness” model was 92.6% (Table 7).\u003c/p\u003e\n\u003cp\u003eTwo additional models were created for the prediction of crunchiness and crackliness, in the same manner as the crispness model but using crunchiness and crackliness ratings for labeling. The overall performance of the crunchiness model was 94.5% (Table 8), while the overall performance of the crackliness model was 96.6% (Tables 9).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 7\u003c/strong\u003e Prediction “crispness” ratings based on sound by machine learning models\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"564\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAcc: 92.58%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrisp 11.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrisp 10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrisp 9.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrisp 8.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNoise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUncertain\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrisp 11.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e90.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e7.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrisp 10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e93.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrisp 9.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e90.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrisp 8.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e90.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNoise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e97.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e95.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1 Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 8\u003c/strong\u003e Prediction of “crunchiness” ratings based on sound by machine learning models\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"583\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAcc: 94.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunch 6.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunch 11.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunch 9.36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunch 9.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNoise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUncertain\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunch 6.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e97.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunch 11.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e95.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunch 9.36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e93.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrunch 9.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e91.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNoise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e97.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e93.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1 Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 9\u003c/strong\u003e Prediction of “crackliness” ratings based on sound by machine learning models\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"583\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAcc: 96.58%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrack 6.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrack 10.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrack 9.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrack 9.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNoise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUncertain\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrack 6.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e96.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrack 10.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrack 9.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCrack 9.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e92.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNoise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e96.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e96.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1 Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThere have been a few studies on the use of neural networks (NN) to assess food texture. Liu \u0026amp; Tan (1999) used specially designed pliers to crush several dry snacks, at two moisture levels, while recording sounds, and assessed crispness only on an unstructured scale from 0 to 10. They analyzed the sounds by \u003cem\u003esignal value dependency,\u0026nbsp;\u003c/em\u003ea frequency-based type of autocorrelation along with \u003cem\u003epower value dependency\u003c/em\u003e, which plots how frequency spectra play out over time. PCA analysis showed 32 features survived screening although these were not explicitly stated. The neural network model was able to classify 10 samples into 4 crispness groups. Okada et al. (2016) used a tooth-like sensor that could detect load and vibrations, and developed an NN model that could classify snacks into several categories. Kato et al. (2019) simultaneously measured sound and load. The results were segmented into average loads in 5 time periods, and integration of frequency spectra in 5 frequency bands. Interestingly, they used convolution neural networks to analyze images of the spectra. They were able to estimate crispness and crunchiness on a scale of 0 to 1, but it is not clear how the scale was developed. Przybl et al. (2020) recorded the acoustic wave disturbance of dried strawberries rolling down a pipe. They found the optimal NN was a Multi-Layer Perceptron network that could classify firmness as measured by a texture analyzer.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; It might be pointed out that each approach for using sound to assess food texture has unique aspects in terms of how the measurement is made, what universe of foods are tested, what signal features are extracted for analysis and what type of neural network architecture is implemented. In our work we emphasized the simultaneous determination of crisp, crunchy and crackly attributes using a small, rounded probe that could fracture and continue through the sample. The sounds were further processed using MFCCs that have proven successful extracting features best suited for mimicking human hearing. In addition, this work used a trained-panel with a structured scale using specified standards for each texture attribute.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;The goal of creating an ML model that could predict the type of chip based on the sound of chip breaking was realized. Furthermore, the model was extended to use sensorial crispness, crunchiness or crackliness ratings from trained panelists and was successfully able to identify how the chips were categorized as to levels of these attributes. However, at this point we cannot claim that the model could predict an intermediate level of crispness, crunchiness or crackliness. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMechanical and acoustical tests were performed on the chip snacks, also yielding data related to crispness, crunchiness and crackliness. Evaluations consistently showed the crispness levels were significantly distinct between each chip type. The data suggest that attributes such as peak force, linear distance or sound intensity might well predict levels of crunchiness or crackliness. Thus, it is still reasonable to conclude that these perceptions are related to the multiple fractures sensed in the mouth as well as the noisy sounds that are produced simultaneously. Crispness, however, was not readily predicted by combinations of physical attributes and this is likely due to panelists switching to crunchiness or crackliness as the relevant nomenclature as the value of these attributes increased. In this context, the use of machine learning models was valuable as they did very well at predicting the chip type or level of sensory attribute for all categories.\u003c/p\u003e\n"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the conception and design of the study. Toby Serrano: data collection, data curation, formal analyses, original draft. William Kerr- supervision and project administration, experimental design, data analysis, review and editing.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eCompeting interest\u003c/strong\u003es\u003c/p\u003e\n\u003cp\u003eThe authors declare there are no financial or non-financial interests that are directly or indirectly related to the work submitted for publication. \u0026nbsp;\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAbdul ZK, Al-Talabani AK (2022) Mel frequency cepstral coefficient and its applications: a review. IEEE Access 10: 122144-122153. https://doi.org/10.1109/ACCESS.2022.3223444\u003c/li\u003e\n\u003cli\u003eAleixandre A, Benavent-Gil Y, Velickova E, Rosell CM (2021) Mastication of crisp bread: role of bread texture and structure on texture perception. Food Res Int 147: 110477. https://doi.org/10.1016/j.foodres.2021.110477\u003c/li\u003e\n\u003cli\u003eBalemarthy S, Sajjanhar A, Zheng JX (2018) Our practice of using machine learning to recognize species by voice. arXiv Prepr. https://doi.org/10.48550/arXiv.1810.09078\u003c/li\u003e\n\u003cli\u003e\u0026Ccedil;arşanba E, Duerrschmid K, Schleining G (2018) Assessment of acoustic-mechanical measurements for crispness of wafer products. J Food Eng 229: 93-101. https://doi.org/10.1016/j.jfoodeng.2017.11.006\u003c/li\u003e\n\u003cli\u003eCastro-Prada EM, Primo-Mart\u0026iacute;n C, Meinders MBJ, Hamer RJ, Van Vliet T (2009) Relationship between water activity, deformation speed, and crispness characterization. J Texture Stud 40: 152-171. https://doi.org/10.1111/j.1745-4603.2009.00173.x\u003c/li\u003e\n\u003cli\u003eChanvrier H, Jakubczyk E, Gondek E, Gumy JC (2014) Insights into the texture of extruded cereals: structure and acoustic properties. Innovative Food Sci Emerg Tech 24: 61-68. https://doi.org/10.1016/j.ifset.2013.11.013\u003c/li\u003e\n\u003cli\u003eChaunier L, Courcoux P, Della Valle GUY, Lourdin D (2005) Physical and sensory evaluation of cornflakes crispness: a multidisciplinary approach. J Texture Stud 36: 93-118. https://doi.org/10.1111/j.1745-4603.2005.00007.x\u003c/li\u003e\n\u003cli\u003eChauvin MA, Younce F, Ross C, Swanson B (2008) Standard scales for crispness, crackliness and crunchiness in dry and wet foods: relationship with acoustical determinations. J Texture Stud 39: 345-368. https://doi.org/10.1111/j.1745-4603.2008.00147.x\u003c/li\u003e\n\u003cli\u003eCheely AN, Pegg RB, Kerr WL, Swanson RB, Huang G, Parrish DR, Kerrihard AL (2018) Modeling sensory and instrumental texture changes of dry-roasted almonds under different storage conditions. LWT-Food Sci Tech 91: 498-504. https://doi.org/10.1016/j.lwt.2018.01.069\u003c/li\u003e\n\u003cli\u003eChen J, Karlsson C, Povey M (2005) Acoustic envelope detector for crispness assessment of biscuits. J Texture Stud 36: 139-156. https://doi.org/10.1111/j.1745-4603.2005.00008.x\u003c/li\u003e\n\u003cli\u003eDacremont C (1995) Spectral composition of eating sounds generated by crispy, crunchy and crackly foods. J Texture Stud 26: 27-43. https://doi.org/10.1111/j.1745-4603.1995.tb00782.x\u003c/li\u003e\n\u003cli\u003eDas N, Mondal A, Chaki J, Padhy N, Dey N (2020) Machine learning models for bird species recognition based on vocalization: a succinct review. Inf Tech Intel Transp Sys 323: 117-124. https://doi.org/10.3233/FAIA200052\u003c/li\u003e\n\u003cli\u003ede Moraes JO, Andreani P, Murta BH, Link JV, Tribuzi G, Laurindo JB, Carciofi BA (2022) Mechanical-acoustical measurements to assess the crispness of dehydrated bananas at different water activities. LWT-Food Sci Tech 154: 112822. https://doi.org/10.1016/j.lwt.2021.112822\u003c/li\u003e\n\u003cli\u003eHymel S (2022) Keyword spotting with edge impulse. GitHub, https://github.com/ShawnHymel/ei-keyword-spotting/tree/master. Accessed 10 February 2025\u003c/li\u003e\n\u003cli\u003eJiang H (2021) Machine Learning Fundamentals: A Concise Introduction. Cambridge University Press\u003c/li\u003e\n\u003cli\u003eKato S, Wada N, Ito R, Shiozaki T, Nishiyama Y, Kagawa T (2019) Snack texture estimation system using a simple equipment and neural network model. Fut Internet 11: 68. https://doi.org/10.3390/fi11030068\u003c/li\u003e\n\u003cli\u003eLink JV, Tribuzi G, de Moraes JO, Laurindo JB (2018) Assessment of texture and storage conditions of mangoes slices dried by a conductive multi-flash process. J Food Eng 239: 8-14. https://doi.org/10.1016/j.jfoodeng.2018.06.024\u003c/li\u003e\n\u003cli\u003eLiu X, Tan J (1999) Acoustic wave analysis for food crispness evaluation. J Texture Stud 30: 397=408. https://doi.org/10.1111/j.1745-4603.1999.tb00227.x\u003c/li\u003e\n\u003cli\u003eLyons J (2009) Mel frequency cepstral coefficient (MFCC) tutorial. Practical cryptography, http://practicalcryptography.com/miscellaneous/machine-learning/guide-mel-frequency-cepstral-coefficients-mfccs/. Accessed 22 April 2025\u003c/li\u003e\n\u003cli\u003eNimbarte N, Khan H, Sendre MD, Ramteke K, Wairagade S (2023) New born baby cry analysis and classification. 4\u003csup\u003eth\u003c/sup\u003e Int Conf Emerging Tech (INCET). 2023: 1-5. https://doi.org/10.1109/INCET57972.2023.10170511\u003c/li\u003e\n\u003cli\u003eOkada S, Nakamoto H, Kobayashi F, Kojima F (2016) Intel Robotics App (ICIRA 2016). In: Kubota N, Kiguchi K, Liu H, Obo T (eds) Springer Cham, Switzerland, https://doi.org/10.1007/978-3-319-43506-0_21\u003c/li\u003e\n\u003cli\u003ePrabhakar H, Kerr WL, Bock CH, Kong F (2023) Effect of relative humidity, storage days, and packaging on pecan kernel texture: Analyses and modeling. J Texture Stud 54: 115-126. https://doi.org/10.1111/jtxs.12723\u003c/li\u003e\n\u003cli\u003ePrimo-Mart\u0026iacute;n C, De Beukelaer H, Hamer RJ, Van Vliet T (2008) Fracture behaviour of bread crust: Effect of bread cooling conditions. J Food Eng 89: 285-290. https://doi.org/10.1016/j.jfoodeng.2008.05.005\u003c/li\u003e\n\u003cli\u003ePrzybył K, Duda A, Koszela K, Stangierski J, Polarczyk M, Gierz L (2020) Classification of dried strawberry by the analysis of the acoustic sound with artificial neural networks. Sensors 20: 499. https://doi.org/10.3390/s20020499\u003c/li\u003e\n\u003cli\u003eRodriguez CR, Angeles D, Chafloque R, Kaseng F, Pandey B (2020) Deep learning audio spectrograms processing to the early COVID-19 detection. 12th Int Conf Comp Intel Comm Netw. 2020: 429-434. https://doi.org/10.1109/CICN49253.2020.9242583\u003c/li\u003e\n\u003cli\u003eRoudaut G, Dacremont C, Meste ML (1998) Influence of water on the crispness of cereal-based foods: acoustic, mechanical, and sensory studies. J Texture Stud 29: 199-213. https://doi.org/10.1111/j.1745-4603.1998.tb00164.x\u003c/li\u003e\n\u003cli\u003eRoudaut G, Dacremont C, P\u0026agrave;mies BV, Colas B, Le Meste M (2002) Crispness: a critical review on sensory and material science approaches. Trends Food Sci 13: 217-227. https://doi.org/10.1016/S0924-2244(02)00139-5\u003c/li\u003e\n\u003cli\u003eSalvador A, Varela P, Sanz T, Fiszman SM (2009) Understanding potato chips crispy texture by simultaneous fracture and acoustic measurements, and sensory analysis. LWT-Food Sci Tech 42: 763-767. https://doi.org/10.1016/j.lwt.2008.09.016\u003c/li\u003e\n\u003cli\u003eStone H, Bleibaum RN, Thomas HA (2020) Sensory Evaluation Practices. Academic Press, New York\u003c/li\u003e\n\u003cli\u003eSuwonsichon T, Peleg M (1998) Instrumental and sensory detection of simultaneous brittleness loss and moisture toughening in three puffed cereals. J Texture Stud 29: 255-274. https://doi.org/10.1111/j.1745-4603.1998.tb00169.x\u003c/li\u003e\n\u003cli\u003eSzczesniak AS, Kahn EL (1971) Consumer awareness of and attitudes to food texture: I. Adults. J Texture Stud 2: 280-295. https://doi.org/10.1111/j.1745-4603.1971.tb01005.x\u003c/li\u003e\n\u003cli\u003eSzczesniak AS, Kahn EL (1984) Texture contrasts and combinations: a valued consumer attribute. J Texture Stud 15: 285-301. https://doi.org/10.1111/j.1745-4603.1984.tb00385.x\u003c/li\u003e\n\u003cli\u003eTaniwaki M, Kohyama K (2012) Mechanical and acoustic evaluation of potato chip crispness using a versatile texture analyzer. J Food Eng 112: 268-273. https://doi.org/10.1016/j.jfoodeng.2012.05.015\u003c/li\u003e\n\u003cli\u003eTrivedi K, Shroff H (2021) identification of deadliest mosquitoes using wing beats sound classification on tiny embedded system using machine learning and Edge Impulse platform. ITU Kaleidosc Conn Phys Virt Worlds. 2021: 1-6. https://doi.org/10.23919/ITUK53220.2021.9662116\u003c/li\u003e\n\u003cli\u003eTunick MH, Onwulata CI, Thomas AE, Phillips JG, Mukhopadhyay S, Sheen S, Cooke PH (2013) Critical evaluation of crispy and crunchy textures: a review. Int J Food Props 16: 949-963. https://doi.org/10.1080/10942912.2011.573116\u003c/li\u003e\n\u003cli\u003eVarela P, Chen J, Fiszman S, Povey MJW (2006) Crispness assessment of roasted almonds by an integrated approach to texture description: texture, acoustics, sensory and structure. J Chemometrics 20: 311-320. https://doi.org/10.1002/cem.1029\u003c/li\u003e\n\u003cli\u003eVicker Z M (1984a) Crackliness: relationships of auditory judgments to tactile judgments and instrumental acoustical measurements. J Food Tex 15: 49-58.\u003c/li\u003e\n\u003cli\u003eVickers ZM (1984b) Crispness and crunchiness\u0026mdash;a difference in pitch? J Texture Stud 15: 157-163. https://doi.org/10.1111/j.1745-4603.1984.tb00375.x\u003c/li\u003e\n\u003cli\u003eVincent JF (1998) The quantification of crispness. J Sci Food Agr 78: 162-168. https://doi.org/10.1002/(SICI)1097-0010(199810)78:2\u0026lt;162::AID-JSFA97\u0026gt;3.0.CO;2-3\u003c/li\u003e\n\u003cli\u003eXu S, Kerr WL (2012) Comparative study of physical and sensory properties of corn chips made by continuous vacuum drying and deep fat frying. LWT-Food Sci Tech 48: 96=101. https://doi.org/10.1016/j.lwt.2012.02.019\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Crispness, crunchiness, AI models, acoustic measurements","lastPublishedDoi":"10.21203/rs.3.rs-6514631/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6514631/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Crispness, crunchiness and crackliness attributes of four snack chips (Lays Classic, Kettle-Cooked, Ruffles and Doritos) was assessed through mechanical and acoustic signatures, coupled with sensory analysis by a trained panel. The highest crispness levels were found for Lay’s Classic chips and greatest crunchiness for Kettle Cooked chips. However, any chip had simultaneous perceptions of crispness, crunchiness and crackliness. For mechanical properties, crunchiness and crackliness were most closely associated with peak force and “linear distance”; for sound they were most closely related to sound intensity and the amount of acoustical energy below 1.7 kHz. A machine learning (ML) model was built based on 200 sound files from each chip, augmented with noise to supply 1000 files for each type. The ML model considered two additional categories of “noise” and “unknown” sounds, and used 80% of the datasets for training and 20% for testing. The prediction accuracy ranged from 89.5% for Doritos to 98% for Kettle-Cooked chips. In addition, the accuracy ranged from 90.0-97.0% for predicting crispness, 91.0-97.0% for crunchiness and 92.5-99.0% for crackliness.","manuscriptTitle":"Machine Learning Models of Sounds to Predict Textural Attributes of Snack Chips","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-05 09:12:17","doi":"10.21203/rs.3.rs-6514631/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"61871069-2e60-44b1-8836-2603c569b83d","owner":[],"postedDate":"May 5th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-25T10:09:04+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-05 09:12:17","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6514631","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6514631","identity":"rs-6514631","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}