Pathological Voices Research Articles

Adversarial Continual Learning to Transfer Self-Supervised Speech Representations for Voice Pathology Detection

Adversarial Continual Learning to Transfer Self-Supervised Speech Representations for Voice Pathology Detection

Promene nivoa polnih hormona sa starenjem i njihov uticaj na glas

The aim of this study is to review modern literature to provide an insight into contemporary research into the etiology, pathophysiology and symptomatology of changes in hormone levels associated with aging and their pattern-consequential relationship with voice quality, with special reference to the analysis of different types of mutations (mutatio precox and mutatio perversa), menstrual and climacteric dysphonia. The criteria for inclusion of works obtained from the search were: works published during the last 15 years, published in English. Compared to the male voice, the female voice does not go through drastic changes during puberty. In males, the increased levels of testosterone and dihydrotestosterone during puberty are responsible for the enlargement of the laryngeal cartilage, which causes the basal tone to decrease by a third of an octave. If the child enters puberty prematurely, this hormone imbalance will lead to the appearance of voice pathology -mutation precox. Mutatio perversa occurs suddenly in females and is characterized by a deep male voice and other signs of virilization. The appearance of submucosal bleeding in the period before menstruation and the increased secretion of estrogen that leads to the appearance of edema will cause premenstrual dysphonia. Voice quality in this period is characterized by vocal fatigue, reduced voice range with loss of high tones, loss of vocal power and hoarseness of voice. Menopausal vocal syndrome is characterized by reduced flexibility and a reduced range of vibrational amplitude. It can be concluded that the vocal cords are a target organ, sensitive to fluctuations in the level of sex hormones due to the physiological aging process and the resulting endocrine disorders during life.

Artificial Intelligence for Clinical Research in Voice Disease

Diagnosis using voice is non-invasive and can be implemented through various voice recording devices; therefore, it can be used as a screening or diagnostic assistant tool for laryngeal voice disease to help clinicians. The development of artificial intelligence algorithms, such as machine learning, led by the latest deep learning technology, began with a binary classification that distinguishes normal and pathological voices; consequently, it has contributed in improving the accuracy of multi-classification to classify various types of pathological voices. However, no conclusions that can be applied in the clinical field have yet been achieved. Most studies on pathological speech classification using speech have used the continuous short vowel /ah/, which is relatively easier than using continuous or running speech. However, continuous speech has the potential to derive more accurate results as additional information can be obtained from the change in the voice signal over time. In this review, explanations of terms related to artificial intelligence research, and the latest trends in machine learning and deep learning algorithms are reviewed; furthermore, the latest research results and limitations are introduced to provide future directions for researchers.

Objective Evaluation of the Pathological Voice Based on Deep Learning Neural Networks in an Algerian hospital environment

In this study, we propose a method based on Recurrent Neural Networks, to objectively evaluate the process of rehabilitation of the pathological voice, in an Algerian clinical environment. We choose Unilateral Laryngeal Paralysis as the pathology of the voice. In this paper, we used a Deep Learning system of pathological voice detection by Long Short Term Memory neural model (LSTM). As the dysphony studied in our work concerns essentially the laryngeal vibration, we choose the acoustic parameters based on the instability of the frequency and the amplitude of the laryngeal vibration: Jitter and Shimmer, Noise parameters and Cepstraux MFCC coefficients (Mel Frequency Cepstral Coefficients). A pathological voice detection rate of 88.65% shows important results brought by the rehabilitation technique adopted in Algerian clinical setting. The exclusive and abusive use of hearing to evaluate the effect of speech rehabilitation in the Algerian hospital environment remains insufficient. It is important to correlate perceptual data with objective methods based on detection and classification methods by introducing relevant acoustic parameters, for an effective and objective management of vocal pathology assessment.

PROSPECTIVE ARABIC LANGUAGE DATA BASE FOR IDENTIFICATION AND RECOGNITION OF VOICE DISORDER


 
 
 Speech is the underlying strong characteristic and voice is the vibration that the air creates as it's pushed out of the lungs and going into the vocal cords. Vocal cords are two folds of tissue within the larynx, sometimes called a speech box, the sound of these strings is what gives rise to speech.Anything that interferes with vocal cord movement or contact can cause a voice disorder. Speech production may often be compromised as social stressors contribute to chronic aphonia or dysphonia.Machine learning algorithm and non-invasive systems may play a major role in the early detection and in tracking and even development of efficient pathological speech diagnosis, based on a computerized acoustic analysis. A strong voice impairment dataset will help in recognition and classification ofincreasing number of voice disorders in and outside the Arab area. The database protocols for AVPD (Arabic voice pathology database) has been developed in a way to prevent previous data base deficiencies in widely used databases like MEEI (Massachusetts eye and ear infirmary) and SVD (Saarbrücken voice database).
 
 

A Cepstral Analysis of Pathological Voice Quality in the Korean Population using Praat

This study aimed to investigate the reference values for cepstral peak prominence (CPP) and smoothed CPP (CPPS) measured using Praat in Korean speakers with the normal, healthy and pathological voice. A total of 4,524 Korean participants with vocally healthy (n=410) and dysphonic voices (n=4,114) participated in this study. The speech task consisted of a sustained vowel /a/ and a sentence reading the Korean passage "Walk". CPP and CPPS values were quickly and automatically measured in sustained vowel and continuous speech tasks using Praat script. Furthermore, three veteran speech language pathologists (SLPs) scored the severity of dysphonia using the GRBAS scale (grade, roughness, breathiness, asthenia, strain) and Consensus Auditory Perceptual Evaluation of Voice (CAPE-V). Three SLPs showed high inter- and intra-rater reliabilities (IRR) in auditory-perceptual (A-P) evaluation. Significant differences were confirmed in CPP and CPPS between the normally healthy and pathological voice groups for both voice tasks (P < 0.01). The measured values of CPP and CPPS varied depending on the laryngeal pathology. In the receiver operating characteristic (ROC) curve analysis, the CPP_Vowel (CPP_V), CPPS_V, CPP_Sentence (CPP_S), and CPPS_S cut-off values were <21.5, <12.0, <19.7, and <10.1, respectively. Through ROC curve analysis, it was confirmed that CPP and CPPS had excellent diagnostic accuracy in distinguishing disordered voice (area under the ROC: 0.951-0.966). We investigated the reference values for CPP and CPPS measured with Praat for Korean speakers and confirmed that cepstral analysis is a promising tool for differentiating pathological voice.

Pathological Voice Detection and Classification Based on Multimodal Transmission Network

ObjectivesDescribing pronunciation features from multiple perspectives can help doctors accurately diagnose the pathological type of a patient's voice. According to the two modal information of sound signal and electroglottography (EGG) signal, this paper proposes a pathological voice detection and classification algorithm based on multimodal transmission network. MethodsFirstly, we used the short-time Fourier transform (STFT) to map the features of the two signals, and designed the Mel filter to obtain the Mel spectogram. Then, the constructed multimodal transmission network extracted features from Mel spectogram and applied Multimodal Transfer Module (MMTM) module. Finally, the fusion layer can integrate multimodal information, and the full connection layer diagnoses and classifies voice pathology according to the fused features. ResultsThe experiment was based on 1179 subjects in Saarbrücken voice database (SVD), and the average accuracy, recall, specificity and F1 score of pathological voice classification reached 98.02%, 98.23%, 97.82% and 97.95% respectively. Compared with other algorithms, the classification accuracy is significantly improved. ConclusionsThe proposed model can integrate multiple modal information to obtain more comprehensive and stable voice features and improve the accuracy of pathological voice classification. Future research will further explore in reducing the time-consuming and complexity of the model.

Exploring the mechanics of fundamental frequency variation during phonation onset.

Fundamental frequency patterns during phonation onset have received renewed interest due to their promising application in objective classification of normal and pathological voices. However, the associated underlying mechanisms producing the wide array of patterns observed in different phonetic contexts are not yet fully understood. Herein, we employ theoretical and numerical analyses in an effort to elucidate the potential mechanisms driving opposing frequency patterns for initial/isolated vowels versus vowels preceded by voiceless consonants. Utilizing deterministic lumped-mass oscillator models of the vocal folds, we systematically explore the roles of collision and muscle activation in the dynamics of phonation onset. We find that an increasing trend in fundamental frequency, as observed for initial/isolated vowels, arises naturally through a progressive increase in system stiffness as collision intensifies as onset progresses, without the need for time-varying vocal fold tension or changes in aerodynamic loading. In contrast, reduction in cricothyroid muscle activation during onset is required to generate the decrease in fundamental frequency observed for vowels preceded by voiceless consonants. For such phonetic contexts, our analysis shows that the magnitude of reduction in the cricothyroid muscle activation and the activation level of the thyroarytenoid muscle are potential factors underlying observed differences in (relative) fundamental frequency between speakers with healthy and hyperfunctional voices. This work highlights the roles of sometimes competing laryngeal factors in producing the complex array of observed fundamental frequency patterns during phonation onset.

The Role of Data Analytics in the Assessment of Pathological Speech—A Critical Appraisal

Pathological voice characterization has received increasing attention over the last 20 years. Hundreds of studies have been published showing inventive approaches with very promising findings. Nevertheless, methodological issues might hamper performance assessment trustworthiness. This study reviews some critical aspects regarding data collection and processing, machine learning-oriented methods, and grounding analytical approaches, with a view to embedding developed clinical decision support tools into the diagnosis decision-making process. A set of 26 relevant studies published since 2010 was selected through critical selection criteria and evaluated. The model-driven (MD) or data-driven (DD) character of the selected approaches is deeply examined considering novelty, originality, statistical robustness, trustworthiness, and clinical relevance. It has been found that before 2020 most of the works examined were more aligned with MD approaches, whereas over the last two years a balanced proportion of DD and MD-based studies was found. A total of 15 studies presented MD characters, whereas seven were mainly DD-oriented, and four shared both profiles. Fifteen studies showed exploratory or prospective advanced statistical analysis. Eighteen included some statistical validation to avail claims. Twenty-two reported original work, whereas the remaining four were systematic reviews of others’ work. Clinical relevance and acceptability by voice specialists were found in 14 out of the 26 works commented on. Methodological issues such as detection and classification performance, training and generalization capability, explainability, preservation of semantic load, clinical acceptance, robustness, and development expenses have been identified as major issues in applying machine learning to clinical support systems. Other important aspects to be taken into consideration are trustworthiness, gender-balance issues, and statistical relevance.

The acoustic voice quality index, version 2.06 and 3.01, for the Persian-speaking population

IntroductionDysphonia assessment includes approaches like acoustic analysis, which is non-invasive and easy to use and provides an understandable numerical output. The Acoustic Voice Quality Index (AVQI) is an acoustic model that can detect dysphonia. The Persian language is spoken by around 70,000,000 native speakers. Since AVQI versions 2.06 and 3.01 have not been validated for the Persian yet, this study investigated their concurrent validity and diagnostic accuracy among the Persian-speaking population. MethodsThis scale development study was conducted from 2020 to 2021 on 180 normophonic and dysphonic native Persian-speaking residents of Mashhad, Iran. Five raters rated the samples by auditory-perceptual-judgments, including Grade from the Grade-Rough-Breathy-Asthenic-Strained (an ordinal scale) and the overall dysphonia severity from the Persian version Consensus Auditory Perceptual Evaluation of Voice (a continuous scale) to investigate both versions’ concurrent validity. The intra- and inter-rater reliability and concurrent validity were evaluated for both scales. Both versions’ diagnostic accuracy was assessed by the receiver operating characteristic, and the optimal thresholds were determined. ResultsAVQI-version-2-Persian thresholds of 3.47 and 4.04 provided sensitivity of 88.30% and 85.53% and specificity of 79.07% and 85.58% by the ordinal and continuous scales, respectively. AVQI-version-3-Persian thresholds of 3.07 and 3.03 also rendered sensitivity of 74.47% and 85.53%, and specificity of 97.67% and 91.35% by the ordinal and continuous scales sequentially. ConclusionThe significant values of concurrent validities and diagnostic accuracies of both versions of AVQI-Persian confirmed that it can discriminate between normal and pathological voices among the Persian-speaking population. Hence, it can be used for screening or diagnosis purposes.

Comparison of one-dimensional and three-dimensional glottal flow models in left-right asymmetric vocal fold conditions.

While the glottal flow is often simplified as one-dimensional (1D) in computational models of phonation to reduce computational costs, the 1D flow model has not been validated in left-right asymmetric vocal fold conditions, as often occur in both normal and pathological voice production. In this study, we performed three-dimensional (3D) and 1D flow simulations coupled to a two-mass model of adult male vocal folds and compared voice production at different degrees of left-right stiffness asymmetry. The flow and acoustic fields in 3D were obtained by solving the compressible Navier-Stokes equations using the volume penalization method with the moving vocal fold wall as an immersed boundary. Despite differences in the predicted flow pressure on vocal fold surface between the 1D and 3D flow models, the results showed reasonable agreement in vocal fold vibration patterns and selected voice outcome measures between the 1D and 3D models for the range of left-right asymmetric conditions investigated. This indicates that vocal fold properties play a larger role than the glottal flow in determining the overall pattern of vocal fold vibration and the produced voice, and the 1D flow simplification is sufficient in modeling phonation, at least for the simplified glottal geometry of this study.

Automatic GRBAS Scoring of Pathological Voices using Deep Learning and a Small Set of Labeled Voice Data

Auditory-perceptual evaluation frameworks, such as the grade-roughness-breathiness-asthenia-strain (GRBAS) scale, are the gold standard for the quantitative evaluation of pathological voice quality. However, the evaluation is subjective; thus, the ratings lack reproducibility due to inter- and intra-rater variation. Prior researchers have proposed deep-learning-based automatic GRBAS score estimation to address this problem. However, these methods require large amounts of labeled voice data. Therefore, this study investigates the potential of automatic GRBAS estimation using deep learning with smaller amounts of data. A dataset consisting of 300 pathological sustained /a/ vowel samples was created and rated by eight experts (200 for training, 50 for validation, and 50 for testing). A neural network model that predicts the probability distribution of GRBAS scores from an onset-to-offset waveform was proposed. Random speed perturbation, random crop, and frequency masking were investigated as data augmentation techniques, and power, instantaneous frequency, and group delay were investigated as time-frequency representations. Five-fold cross-validation was conducted, and the automatic scoring performance was evaluated using the quadratic weighted Cohen's kappa. The results showed that the kappa values of the automatic scoring performance were comparable to those of the inter-rater reliability of experts for all GRBAS items and the intra-rater reliability of experts for items G, B, A, and S. Random speed perturbation was the most effective data augmentation technique overall. When data augmentation was applied, power was the most effective for items G, R, A, and S; for Item B, combining group delay and power yielded additional performance gains. The automatic GRBAS scoring achieved by the proposed model using scant labeled data was comparable to that of experts. This suggests that the challenges resulting from insufficient data can be alleviated. The findings of this study can also contribute to performance improvements in other tasks such as automatic voice disorder detection.

An Intelligent System for the Diagnosis of Voice Pathology Based on Adversarial Pathological Response (APR) Net Deep Learning Model

The work investigates the use of two types of glottal flow derivative-based image variants of the input signal with an n-dilated (nD)-inception-layers-based deep learning model for providing optimal labels. The authors have proposed an n-dilated (nD) inception layer-based adversarial pathological response (APR) net deep learning model. This model is trained using the two image databases separately in an adversarial manner so that when a test image is common to test image is applied to both the networks. The results show a mean accuracy of 96.82%, 96.36%, and 99.35% for the Glottal inverse filtering with extended Kalman Filter-Morse scalogram (GIFEKF-MS) APRNet, Glottal inverse filtering with extended Kalman Filter-spectrogram (GIFEKF-S) APRNet, and proposed APR fusion net respectively using the VOice ICar fEDerico II (VOICED) dataset; and mean accuracies 95.67%, 93.27%, and 99.04% for the GIFEKF-MS APRNet, GIFEKF-S APRNet, and proposed APR fusion net respectively using the Saarbrucken voice database (SVD)dataset.

The Effectiveness of Supervised Machine Learning in Screening and Diagnosing Voice Disorders: Systematic Review and Meta-analysis.

BackgroundWhen investigating voice disorders a series of processes are used when including voice screening and diagnosis. Both methods have limited standardized tests, which are affected by the clinician’s experience and subjective judgment. Machine learning (ML) algorithms have been used as an objective tool in screening or diagnosing voice disorders. However, the effectiveness of ML algorithms in assessing and diagnosing voice disorders has not received sufficient scholarly attention.ObjectiveThis systematic review aimed to assess the effectiveness of ML algorithms in screening and diagnosing voice disorders.MethodsAn electronic search was conducted in 5 databases. Studies that examined the performance (accuracy, sensitivity, and specificity) of any ML algorithm in detecting pathological voice samples were included. Two reviewers independently selected the studies, extracted data from the included studies, and assessed the risk of bias. The methodological quality of each study was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 tool via RevMan 5 software (Cochrane Library). The characteristics of studies, population, and index tests were extracted, and meta-analyses were conducted to pool the accuracy, sensitivity, and specificity of ML techniques. The issue of heterogeneity was addressed by discussing possible sources and excluding studies when necessary.ResultsOf the 1409 records retrieved, 13 studies and 4079 participants were included in this review. A total of 13 ML techniques were used in the included studies, with the most common technique being least squares support vector machine. The pooled accuracy, sensitivity, and specificity of ML techniques in screening voice disorders were 93%, 96%, and 93%, respectively. Least squares support vector machine had the highest accuracy (99%), while the K-nearest neighbor algorithm had the highest sensitivity (98%) and specificity (98%). Quadric discriminant analysis achieved the lowest accuracy (91%), sensitivity (89%), and specificity (89%).ConclusionsML showed promising findings in the screening of voice disorders. However, the findings were not conclusive in diagnosing voice disorders owing to the limited number of studies that used ML for diagnostic purposes; thus, more investigations are needed. While it might not be possible to use ML alone as a substitute for current diagnostic tools, it may be used as a decision support tool for clinicians to assess their patients, which could improve the management process for assessment.Trial RegistrationPROSPERO CRD42020214438; https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=214438

Pathological Voice Detection Based on Phase Reconstitution and Convolutional Neural Network

The nonlinear dynamic features can effectively describe the acoustic characteristics of normal and pathological voice. In this paper, the phase space reconstruction and convolution neural network are used to classify the normal and pathological voice. The phase space information of normal and pathological voice is reconstructed using delay time and embedding dimension, the one-dimensional signal is converted to a two-dimensional matrix, and the reconstructed trajectory graph sample of the signal is generated. The trajectory graph samples are used as the input of the VGG-like convolutional neural network, and the graphical features are extracted to achieve a classification of normal and pathological voice. In order to overcome the lack of clinical data, a data enhancement scheme is used. The experiment which classifies the normal and pathological voice is carried out on three pathological databases respectively, i.e. the Massachusetts eye and ear infirmary (MEEI) database, Saarbrücken voice database (SVD) database, and a clinical database collected by the authors. Five-fold cross validation is used and the average recognition rates on the three databases are 99.42%, 97.30% and 95.88% respectively. The average recognition rates are 96.04% and 92.27% for normal, vocal fold paralysis and vocal fold non-paralysis voice in MEEI database and SVD database. The experimental results show that the method has high classification recognition rate and good robustness, and has certain universal applicability for the recognition of the normal and pathological voice.

Validity, Reliability and Reproducibility of the “Extended GRBAS Scale,” A Comprehensive Perceptual Evaluation of Dysphonia

<h2>Summary</h2><h3>Objective</h3> To validate the Extended GRBAS Scale (GS) and to investigate its inter- and intra-rater reliability and reproducibility. The Extended GS contains the Standard GS with 8 added perceptual parameters similar to those present in the CAPE-V. <h3>Methods</h3> We have carried out a retrospective study on 480 evaluation forms for the perceptual evaluation of dysphonia with the Extended GS, examined by four experienced raters, 240 evaluations on running speech and 240 on sustained vowel /a/. The four raters examined, random and blind, 90 audio files of pathologic voices and 30 samples of normal voices. The pathologic samples were the pre- and post-phono surgery voices of 15 selected patients. The 30 pre-operative voice samples (15 of running speech and 15 of sustained vowel) were blindly re-assessed in a different order. Each rater annotated the difficulty in rating the added perceptual parameter and its "importance" for the assessment of dysphonia in that patient. Voices were recorded in standard conditions. All audio files were analyzed by the CSL and by the PRAAT, for an acoustic objective evaluation of meanF0 of conversational voice, Jitter%, Shimmer%, NHR, CPPS; we also verified the presence of sub-harmonics and undulation of F0 in the spectrogram. <h3>Results</h3> The Extended GS showed good inter- and intra-rater reliability and reproducibility. It provided a more complete assessment than the Standard GS for documenting voice changes after Phono surgery. The 8 added perceptual parameters were judged "Slightly Important" in 34.4% of the samples, "Very Important" in 32.5% of the samples, "Not Difficult" in 80.4% of the samples. Significant correlations were found between perceptual evaluation with the Extended GS and acoustic analysis. <h3>Conclusions</h3> The Extended GS is a useful tool for the voice clinician, who can use the Standard GS, or the Reduced GS (GRB) for a basic evaluation. The Extended GS provides for a comprehensive perceptual evaluation of dysphonia pre and post treatment.

A novel convolutional neural network based dysphonic voice detection algorithm using chromagram

&lt;span&gt;This paper presents a convolutional neural network (CNN) based non-invasive pathological voice detection algorithm using signal processing approach. The proposed algorithm extracts an acoustic feature, called chromagram, from voice samples and applies this feature to the input of a CNN for classification. The main advantage of chromagram is that it can mimic the way humans perceive pitch in sounds and hence can be considered useful to detect dysphonic voices, as the pitch in the generated sounds varies depending on the pathological conditions. The simulation results show that classification accuracy of 85% can be achieved with the chromagram. A comparison of the performances for the proposed algorithm with those of other related works is also presented.&lt;/span&gt;

Introducing a New Dosimeter for the Assessment and Monitoring of Vocal Risk Situations and Voice Disorders

There are many physiological parameters recorded by devices that are becoming more affordable, precise and accurate. However, the lack of development in the recording of voice parameters from the physiological or medical point of view is striking, given that it is a fundamental tool for the work of many people and given the high incidence and prevalence of voice pathologies that affect people's communication. In this paper we perform a complete literature review on the dosimeters used in voice research and to present a prototype dosimeter with a pilot study to show its capabilities. We conducted a literature review using the keywords [MONITORING], [PHONATION], [ACCUMULATOR], [PORTABLE], [DOSIMETRY], [VOICE] searching in PubMed, Trip Database, HONcode, and SciELO search engines. From our review of dosimeter designs, we created our own prototype consisting of two main components: a Knowles Electronics BU-7135-0000 accelerometer mounted on a neck brace; and the ultra-low power MSP430FR5994 microcontroller. The selected sampling frequency was 2048 Hz. The device calculates the F0 every 250 ms and the amplitude and phonation activity every 31.25 ms. A pilot study was conducted using 2 subjects: one male during 11 days and one female during 14 days. This work includes devices that have been created during the last 45 years as tools for the diagnosis and monitoring of the treatment of cases of vocal pathology and for the detection of phonatory patterns or risk situations for developing voice disorders or vocal pathologies. We also present recordings with our new device on the pattern of daily talk time, the fundamental frequency and the relative intensity of two subjects on different days. Interesting work has been done in the development of voice dosimeters with different approaches. In our experience it is not possible to access them for research and they are not yet in clinical use. It is possible that a joint approach with voice and voice disorders professionals and engineers working closely together could take advantage of current technology to develop a fully portable, useful, and efficient system.

Multi-Scale Recurrence Quantification Measurements for Voice Disorder Detection

Due to the complexity and non-stationarity of the voice generation system, the nonlinearity of speech signals cannot be accurately quantified. Recently, the recurrence quantification analysis method has been used for voice disorder detection. In this paper, multiscale recurrence quantification measures (MRQMs) are proposed. The signals are reconstructed in the high-dimensional phase space at the equivalent rectangular bandwidth scale. Recurrence plots (RPs) combining the characteristics of human auditory perception are drawn with an appropriate recurrence threshold. Based on the above, the nonlinear dynamic recurrence features of the speech signal are quantized from the recurrence plot of each frequency channel. Furthermore, this paper explores the recurrence quantification thresholds that are most suitable for pathological voices. Our results show that the proposed MRQMs with support vector machine (SVM), random forest (RF), Bayesian network (BN) and Local Weighted Learning (LWL) achieve an average accuracy of 99.45%, outperforming traditional features and other complex measurements. In addition, MRQMs also have the potential for multi-classification of voice disorder, achieving an accuracy of 89.05%. This study demonstrates that MRQMs can characterize the recurrence characteristic of pathological voices and effectively detect voice disorders.

Using SincNet for Learning Pathological Voice Disorders

Deep learning techniques such as convolutional neural networks (CNN) have been successfully applied to identify pathological voices. However, the major disadvantage of using these advanced models is the lack of interpretability in explaining the predicted outcomes. This drawback further introduces a bottleneck for promoting the classification or detection of voice-disorder systems, especially in this pandemic period. In this paper, we proposed using a series of learnable sinc functions to replace the very first layer of a commonly used CNN to develop an explainable SincNet system for classifying or detecting pathological voices. The applied sinc filters, a front-end signal processor in SincNet, are critical for constructing the meaningful layer and are directly used to extract the acoustic features for following networks to generate high-level voice information. We conducted our tests on three different Far Eastern Memorial Hospital voice datasets. From our evaluations, the proposed approach achieves the highest 7%–accuracy and 9%–sensitivity improvements from conventional methods and thus demonstrates superior performance in predicting input pathological waveforms of the SincNet system. More importantly, we intended to give possible explanations between the system output and the first-layer extracted speech features based on our evaluated results.