Vowel Samples Research Articles

Can acoustic measurements predict gender perception in the voice?

Purpose To determine if there is an association between vocal gender presentation and the gender and context of the listener. Method Quantitative and transversal study. 47 speakers of Brazilian Portuguese of different genders were recorded. Recordings included sustained vowel emission, connected speech, and the expressive recital of a poem. Subsequently, four scripts were used in Praat to extract 16 acoustic measurements related to prosody. Voices underwent Auditory-Perceptual Assessment (APA) of the gender presentation by 236 people [65 speech and language pathologist (SLP) with experience in the area of the voice (SLP), 101 cisgender people (CG), and 70 transgender and non-binary people (TNB)]. Gender presentation was evaluated by visual analogue scale. Agreement analyses were executed among quantitative variables and multiple linear regression models were generated to predict APA, taking the judge context/gender and speaker gender into consideration. Results Acoustic analysis revealed that cis and transgender women had higher median fundamental frequency (fo) values than other genders. Cisgender women exhibited greater breathiness, while cisgender men showed more vocal quality deviations. In terms of APA, significant differences were observed among judge groups: SLP judged vowel samples differently from other groups, and TNB judged speech samples differently (both p<0.001). The predictive measures for the APA varied based on the sample type, speaker gender, and judge group. For vowel samples, only SLP judges had predictive measures (fo and ABI Jitter) for cisgender speakers. In number counting samples, predictive measures for cisgender speakers included fomed and HNR for CG judges, and fomed for both SLP and TNB judges. For transgender and non-binary speakers, predictive measures were fomed for CG and SLP judges, and fomed, CPPs, and ABI for TNB judges. In the poem recital task, predictive measures for cisgender speakers were fomed and HNR for both SLP and CG judges, with additional measures of cvint and sr for CG judges, and fomed, HNR, cvint, and fopeakwidth for TNB judges. For transgender and non-binary speakers, the predictive measures included a wider range of acoustic features such as fomed, fosd, sr, fomin, emph, HNR, Shimmer, and fo peakwidth for SLP judges, and fomed, fosd, sr, fomax, emph, HNR, and Shimmer for CG judges, while TNB judges considered fomed, sr, emph, fosd, Shimmer, HNR, Jitter, and fomax. Conclusions There is an association between the perception of gender presentation in the voice and the gender or context of the listener and the speaker. Transgender and non-binary judges diverged to a higher degree from cisgender and SLP judges. Compared to the evaluation of cisgender speakers, all judge groups used a greater number of acoustic measurements when analyzing the speech of transgender and non-binary individuals in the poem recital samples.

Validation of the Acoustic Voice Quality Index Version 03.01 in Turkish

ObjectivesThe aim of this study was to validate the Acoustic Voice Quality Index (AVQI) version 3.01 in the Turkish-speaking population. Materials and methodsConcatenated voice samples of the sustained vowel [a:] and continuous speech were collected from 127 dysphonic and 128 normophonic participants. The auditory-perceptual evaluation was performed by five experienced raters using the Grade parameter of the GRBAS scale. Rater reliability, concurrent validity, diagnostic accuracy, and differences between normophonic and dysphonic groups were analyzed for the AVQI version 3.01. ResultsThe number of syllables for the standardized reading text with the concatenation of the voiced parts lasting around 3 seconds (mean=3.84s) was 36. The ICC values of intra-rater reliability of G scores of five raters were excellent (mean ICC=0.934), and of inter-rater reliability they varied between moderate and excellent (mean ICC=0.786). AVQIv3 demonstrated a high diagnostic accuracy with AUC = 0.906 in identifying disrupted versus normal voice quality. With sensitivity of 80% and specificity of 94%, AVQIv3 = 2.345 was the cut-off point that differentiated most accurately between normophonic and dysphonic voices in Turkish. ConclusionAVQIv3 is an ecologically valid tool for objective differentiation between dysphonic and normal voices in the Turkish language.

Effectiveness of Temporal Auditory Skills Training Associated With Conventional Auditory Training in the Auditory-Perceptual Judgment of Voice: Preliminary Data

To verify whether there are differences in the performance and reliability of speech-language-hearing students between the moments of auditory-perceptual judgment (APJ) of voice before and after temporal auditory skills training and conventional auditory-perceptual training (APT). Both trainings were applied to 30 speech-language pathology students, with the following stages: 1-Initial APJ of the general grade (GG), breathiness (GB), and roughness (GR) with 30 samples of the sustained vowel [Ɛ], plus 20% of repetitions (six samples), analyzed with a visual analog scale; 2-Temporal auditory skills training on the online platform Afinando o Cérebro, with a total duration of 4hours; 3-APJ of GG, GR, and GB after training in Afinando o Cérebro; 4-APT in the classroom with a total duration of 6hours; 5-APJ of GG, GR, and GB after APT. There was a significant difference in the students' performance in the APJ of GG after the training. The agreement in identifying voices as dysphonic or nondysphonic increased after each training. Temporal auditory skills training associated with APT has the potential to improve the performance of speech-language-hearing students in APJ of voice, especially in GG.

Cepstral Acoustic Measurements: Influence of Speech Task and Degree of Vocal Deviation

ObjectiveTo analyze whether there are differences in the cepstral measures obtained in different speech tasks, depending on the presence and degree of vocal deviation, and to analyze if there is a correlation between the cepstral measures obtained from different speech tasks and the general degree of vocal deviation. MethodAnalysis of 258 vocal samples of the sustained vowel [a] and connected speech (counting numbers) from a database, including 160 dysphonic and 98 non-dysphonic voices. The counting number sample were edited in three different durations: counting from 1 to 10, from 1 to 11, and from 1 to 20. Five SLP, voice specialists, carried out the perceptual-auditory judgment (PAJ) of the overall degree of vocal deviation (ODD) using the G from the GRBAS scale. We extracted the cepstral peak prominence (CPP) and smoothed cepstral peak prominence (CPPS) measurements from all the vocal samples using an extraction script in the free software Praat. ResultsCPP and CPPS were different between dysphonic and non-dysphonic individuals, regardless of the speech task, with lower values for dysphonic. Also, CPP values between the vowel and the connected speech tasks were different between both groups. Only the CPPS showed differences between all the speech tasks depending on the degree of vocal deviation. There was a strong negative correlation between the CPPSVowel, CPPS10, CPPS11, CPPS20, and the ODD, and a moderate negative correlation between CPPVowel, CPP10, CPP11, CPP20, and ODD. ConclusionThere are differences in the cepstral measures obtained in different speech tasks, depending on the presence of dysphonia and ODD. CPP and CPPS values are different between dysphonic and non-dysphonic individuals in all speech tasks. There is a moderate negative correlation between CCP in the different speech tasks and ODD while there is a strong negative correlation between CPPS in the different speech tasks and ODD.

Evidence-Based Recommendations for Tablet Recordings From the Bridge2AI-Voice Acoustic Experiments

BackgroundAs part of a larger goal to create best practices for voice data collection to fuel voice AI research, the objective of this study was to investigate the ability of readily available iOS and Android tablets with and without low-cost headset microphones to produce recordings and subsequent acoustic measures of voice comparable to “research quality” instrumentation. MethodsRecordings of 24 sustained vowel samples representing a wide range of typical and disordered voices were played via a head-and-torso model and recorded using a research quality standard microphone/preamplifier/audio interface. Acoustic measurements from the standard were compared to two popular tablets using their built-in microphones and with low-cost headset microphones at different distances from the mouth. ResultsVoice measurements obtained via tablets + headset microphones close to the mouth (2.5 and 5cm) strongly correlated (r’s > 0.90) with the research standard and resulted in no significant differences for measures of vocal frequency and perturbation. In contrast, voice measurements obtained using the tablets’ built-in microphones at typical reading distances (30 and 45cm) tended to show substantial variability in measurement, greater mean differences in voice measurements, and relatively poorer correlations vs. the standard. ConclusionFindings from this study support preliminary recommendations from the Bridge2AI-Voice Consortium recommending the use of smartphones paired with low-cost headset microphones as adequate methods of recording for large-scale voice data collection from a variety of clinical and non-clinical settings. Compared to recording using a tablet directly, a headset microphone controls for recording distance and reduces the effects of background noise, resulting in decreased variability in recording quality. Data Availability StatementData supporting the results reported in this article may be obtained upon request from the contact author.

Usefulness of Direct Magnitude Estimation (DME) and Acoustic Analysis in Measuring Dysphonia Severity

ObjectivesThe purposes of this study were (1) to analyze the reliability of DME in auditory perceptual assessments measuring dysphonia severity and (2) analyze the relationship between DME and four acoustic parameters (CPP, CPPs, AVQI, and ABI) and (3) predict dysphonia severity based on DME using four acoustic parameters. Study Design161 voice samples for dysphonia patients were used. In this study, we combined sustained vowel samples and connected speech samples using the Praat software to make the concatenated samples for implementing acoustic analysis and auditory perceptual assessments. For acoustic analysis, we analyzed each value of CPP, CPPs, AVQI, and ABI. For auditory perceptual assessments, three speech language pathologists evaluated dysphonia severity from the concatenated samples. Finally, we performed a stepwise multiple regression analysis to verify which combination of the four acoustic parameters could best predict perceived dysphonia severity based on the DME. ResultsDME was found to have high reliability for auditory perceptual assessments measuring dysphonia severity and there was a significant correlation between DME and four acoustic parameters. Finally, a two-variable model (AVQI and ABI) was useful for predicting perceived dysphonia severity based on the DME. ConclusionsWe verified the usefulness of DME scales in judging the dysphonia severity of dysphonic patients when used with acoustic analysis. Also, two-variable model was a useful to predict perceived dysphonia severity based on the DME.

Voice disorder discrimination using vowel acoustic measures in female speakers.

Sustained vowels are important vocal tasks that have been investigated in discriminating voice disorders using acoustic analysis. To date, no study has combined vowel acoustic measures only that evaluate major aspects of the pathological voice signals in voice disorder discrimination. To investigate the value of vowel acoustic measures that quantify glottal noise, signal stability, signal periodicity, spectral slope and overall voice quality in discriminating female speakers with and without voice disorders. Sustained vowel /ɑ/ samples were extracted from 133 voice-disordered female patients and 97 non-voice disordered female speakers and were signal typed prior to analysis. Praat software was used to measure harmonics-to-noise ratio (HNR), glottal-to-noise excitation ratio (GNE), the standard deviation of fundamental frequency (F0SD) and cepstral peak prominence (CPPp); and the Analysis of Dysphonia in Speech and Voice (ADSV) program was used to measure CPPadsv, low/high spectral ratio (LH) and the cepstral/spectral index of dysphonia (CSID). Outcome measures included sensitivity, specificity, and discrimination accuracy. As individual acoustic measures, only spectral-based measures showed good (CPPadsv) and acceptable (CSID) discrimination results. The HNR, GNE and CPPp measures had acceptable sensitivity but poor or non-acceptable specificity and discrimination accuracy. Logistic regression models with all Praat measures (F0SD, HNR, GNE, CPPp) plus ADSV measures (CPPadsv, LH or CSID) provided excellent sensitivity, good-to-excellent specificity and excellent discrimination accuracy. ROC analysis for all individual measures showed that CPPadsv, CSID, CPPp, GNE and F0SD had the highest area under the curve (AUC) values. A combination of acoustic measures that evaluate the major aspects of vocal dysfunction resulted in good to excellent voice discrimination outcomes. Individual acoustic measures had lower discrimination ability than combined measures. The findings implied that acoustic measures extracted from a prolonged vowel were useful in voice disorder discrimination. What is already known on this subject Acoustic measures hold great value in discriminating voice disorders from normal voices. However, no study has evaluated discrimination values of a combination of sustained vowel acoustic measures that quantify additive noise, signal stability, signal periodicity, spectral slope and overall voice quality in single-gender cohorts. Previous studies have not used signal typing (the classification of the acoustic signals) for time-based measures, impacting the reliability of discrimination. What this study adds to the existing knowledge This study was the first to implement signal typing to include sustained vowel samples of Types 1 and 2 signals for discrimination statistics. We showed that a combination of vocal acoustic measures using time- and spectral-based extraction from the sustained /ɑ/ vowel evaluating additive noise, signal stability, signal periodicity, spectral slope and overall voice quality resulted in good to excellent sensitivity, specificity and discrimination accuracy. As individual measures, traditional time-based measures such as HNR had rather limited discrimination values whilst spectral-based measures provided higher discrimination values. Measures that are sensitive to signal types have low discrimination ability. What are the potential or actual clinical implications of this work? The sustained vowel /ɑ/ is a relevant, universal vocal task for clinical application using acoustic measures to discriminate female speakers with and without voice disorders if signal typing is implemented. Clinical voice assessment using vowels may not be effective if relying solely on time-based measurements. Spectral-based measures perform better in voice disorder discrimination given their insensitivity to signal types. The most effective voice disorder discrimination could only be obtained using a combination of acoustic measures that quantify major phenomena in the signals of disordered voices. Using measures extracted from both programs, Praat and ADSV, is useful given that specific settings in a program may impact on discrimination accuracy.

Effect of Face Masks on Voice Quality Associated with Young and Older Chinese Adult Speakers

ObjectivesFace masks have become important after the pandemic, but the change in voice when wearing a face mask is still unclear. The study investigated the effect of face masks on the acoustic and perceptual characteristics of voice quality associated with young and older Chinese adults. MethodsVoice samples of the sustained vowel /a/ and continuous speech produced by 44 older and 61 young adults with and without an ASTM level-3 surgical face mask were recorded and analyzed. Perceptual and acoustic parameters including mean fundamental frequency (F0) and intensity, perturbation measures (jitter and shimmer), harmonic-to-noise ratio (HNR), smoothed cepstral peak prominence (CPPs), and long-term average spectrum (LTAS) measures were obtained and compared. ResultsWhen comparing masked to unmasked voices, for both male and female speakers, F0 and intensity showed no significant changes, except for F0 of continuous speech, which increased significantly. Meanwhile, perturbation measures including jitter and shimmer were reduced, HNR and CPPs were increased, and LTAS measures encompass low frequency mean spectral energy (MSE), high frequency MSE, and spectral tilt (ST) were different. For perceptual measures, the overall grade of dysphonia, and roughness were reduced, except for the breathiness among older male speakers, while the other vocal qualities were not changed. Between young and older speakers, significant differences in shimmer, CPPs, and perceived breathiness among male speakers, and low frequency MSE among female speakers were found. ConclusionWearing a surgical mask appeared to change the perceived voice quality. This is supported by the change in perturbation and LTAS measures, and HNR and CPPs values. In addition, some differences between young and older male adults were observed. Communication effectiveness may be affected when wearing surgical masks due to changes in voice quality. Additionally, clinicians need to exercise caution in evaluating the voice quality of clients when wearing face masks.

Validity of Acoustic Measures Obtained Using Various Recording Methods Including Smartphones With and Without Headset Microphones.

The goal of this study was to assess various recording methods, including combinations of high- versus low-cost microphones, recording interfaces, and smartphones in terms of their ability to produce commonly used time- and spectral-based voice measurements. Twenty-four vowel samples representing a diversity of voice quality deviations and severities from a wide age range of male and female speakers were played via a head-and-thorax model and recorded using a high-cost, research standard GRAS 40AF (GRAS Sound & Vibration) microphone and amplification system. Additional recordings were made using various combinations of headset microphones (AKG C555 L [AKG Acoustics GmbH], Shure SM35-XLR [Shure Incorporated], AVID AE-36 [AVID Products, Inc.]) and audio interfaces (Focusrite Scarlett 2i2 [Focusrite Audio Engineering Ltd.] and PC, Focusrite and smartphone, smartphone via a TRRS adapter), as well as smartphones direct (Apple iPhone 13 Pro, Google Pixel 6) using their built-in microphones. The effect of background noise from four different room conditions was also evaluated. Vowel samples were analyzed for measures of fundamental frequency, perturbation, cepstral peak prominence, and spectral tilt (low vs. high spectral ratio). Results show that a wide variety of recording methods, including smartphones with and without a low-cost headset microphone, can effectively track the wide range of acoustic characteristics in a diverse set of typical and disordered voice samples. Although significant differences in acoustic measures of voice may be observed, the presence of extremely strong correlations (rs > .90) with the recording standard implies a strong linear relationship between the results of different methods that may be used to predict and adjust any observed differences in measurement results. Because handheld smartphone distance and positioning may be highly variable when used in actual clinical recording situations, smartphone + a low-cost headset microphone is recommended as an affordable recording method that controls mouth-to-microphone distance and positioning and allows both hands to be available for manipulation of the smartphone device.

Effects of Semi-occluded Vocal Tract Exercises with Different Vibration Sources and Duration on Healthy Adults’ Voice

Objectives: This study aimed to explore the impact of semi-occluded vocal tract exercises (SOVTE) with varied vibration sources and durations on the vocal performance of healthy adults. In addition, we sought to confirm the safety of voiced high-frequency oscillation (VHFO) intervention and provide basic data for performances in relation to straw phonation (SP).Methods: The study included 34 healthy adults with no vocal impairments, comprising 8 males and 26 females. The mean age of the participants was 23.00±2.06 years. To assess vocal characteristics during VHFO and SP, voice samples were recorded at three specific time points: prior to performing SOVTE (M0), after 1 minute of exercise (M1), and after 3 minutes of exercise (M3). Cepstral, perturbation, electroglottographic analyses were conducted on the recorded samples of vowel and sentences. Psychometric evaluation was conducted using questionnaires to gauge the self-reported severity of voice and vocal fatigue.Results: No statistically significant differences were observed in the average number of performances or duration per phonation between the two methods. When compared to baseline, CPPV significantly increased after 1 minute and 3 minutes of performance and Acoustic Psychometric Severity Index of Dysphonia decreased after 3 minutes of performance in VHFO. Both methods exhibited no significant differences in terms of psychometric and electroglottographic measurements.Conclusion: Results objectively confirmed the safety of VHFO intervention. This study is clinically significant as it lays the groundwork for future development of optimal treatment regimens and therapeutic programs for voice disorder patients, using foundational data on VHFO.

Voice pathology identification system using a deep learning approach based on unique feature selection sets

AbstractVoice pathology diagnosis requires extracting significant features from voice signals, and classical machine learning models can overfit to the training data, which can cause difficult issues and pose challenges. The study aimed to develop a reliable and efficient system for identifying voice pathologies utilizing the long short‐term memory (LSTM) method. The study combined unique feature sets such as the mel frequency cepstral coefficients (MFCCs), zero crossing rate (ZCR), and mel spectrograms, which have not been used together in previous works. Voice pathology identification improved the accuracy rate using the LSTM approach on the Saarbruecken voice database (SVD) samples. The best results achieved by the proposed system showed an accuracy rate of 99.3% for /u/ vowel samples in neutral pitch, 99.2% for /a/ vowel samples in high pitch, 99% for /i/ vowel samples in neutral pitch, and 99.2% for sentence samples. The experimental results were evaluated utilizing accuracy, precision, specificity, sensitivity, and F1 measures. Additionally, the study compared the performance of LSTM with that of artificial neural networks (ANNs) and found that LSTM achieved better outcomes.

Smartphone Recordings are Comparable to “Gold Standard” Recordings for Acoustic Measurements of Voice

PurposeThe purpose of this study was to assess the relationship and comparability of cepstral and spectral measures of voice obtained from a high-cost “flat” microphone and precision sound level meter (SLM) vs. high-end and entry level models of commonly and currently used smartphones (iPhone i12 and iSE; Samsung s21 and s9 smartphones). Device comparisons were also conducted in different settings (sound-treated booth vs. typical “quiet” office room) and at different mouth-to-microphone distances (15 and 30 cm). MethodsThe SLM and smartphone devices were used to record a series of speech and vowel samples from a prerecorded diverse set of 24 speakers representing a wide range of sex, age, fundamental frequency (F0), and voice quality types. Recordings were analyzed for the following measures: smoothed cepstral peak prominence (CPP in dB); the low vs high spectral ratio (L/H Ratio in dB); and the Cepstral Spectral Index of Dysphonia (CSID). ResultsA strong device effect was observed for L/H Ratio (dB) in both vowel and sentence contexts and for CSID in the sentence context. In contrast, device had a weak effect on CPP (dB), regardless of context. Recording distance was observed to have a small-to-moderate effect on measures of CPP and CSID but had a negligible effect on L/H Ratio. With the exception of L/H Ratio in the vowel context, setting was observed to have a strong effect on all three measures. While these aforementioned effects resulted in significant differences between measures obtained with SLM vs. smartphone devices, the intercorrelations of the measurements were extremely strong (r's > 0.90), indicating that all devices were able to capture the range of voice characteristics represented in the voice sample corpus. Regression modeling showed that acoustic measurements obtained from smartphone recordings could be successfully converted to comparable measurements obtained by a "gold standard" (precision SLM recordings conducted in a sound-treated booth at 15 cm) with small degrees of error. ConclusionsThese findings indicate that a variety of commonly available modern smartphones can be used to collect high quality voice recordings usable for informative acoustic analysis. While device, setting, and distance can have significant effects on acoustic measurements, these effects are predictable and can be accounted for using regression modeling.

Analysis of different performance times of the voiced trill technique in older women.

To analyze and compare the immediate vocal effects of the voiced trill technique in the assessment of acoustic and auditory-perceptual measures of older women with and without self-perceived vocal changes. Clinical, quasi-experimental study in older women, aged 60 to 70 years (n=53). A questionnaire on vocal self-perception, voice, and laryngeal assessment was applied, before and after performing the voiced trill technique. Before and during intervals of the technique, sustained vowel samples were collected, totaling four samples. Older women were divided into two groups: one with self-perceived voice changes (n=25), and the other without self-perceived voice changes (n=28). Auditory-perceptual assessments and acoustic analysis were performed. Statistical tests were used to correlate the data: ANOVA Test for repeated measures, Friedman Test, Wilcoxon Test, and Pearson's Chi-Square Test. For all tests, the significance level was set at 5%. There was a predominance of moderate dysphonia in both groups, according to the auditory-perceptual judgment. There was no statistically significant difference between the groups in the assessment of the auditory-perceptual analysis regarding voice changes (improved, worsened, and unaltered voices) before and after the different technique performance times. Most older women improved their voice after 1 minute of performing the technique. Older women often have voice changes when considering the perceptual judgment of the voice. There was no scientific evidence as to the ideal time to obtain a better effect on older women's voices.

Análise dos tempos de execução da técnica de sons vibrantes em idosas

ABSTRACT Purpose To analyze and compare the immediate vocal effects of the voiced trill technique in the assessment of acoustic and auditory-perceptual measures of older women with and without self-perceived vocal changes. Methods Clinical, quasi-experimental study in older women, aged 60 to 70 years (n=53). A questionnaire on vocal self-perception, voice, and laryngeal assessment was applied, before and after performing the voiced trill technique. Before and during intervals of the technique, sustained vowel samples were collected, totaling four samples. Older women were divided into two groups: one with self-perceived voice changes (n=25), and the other without self-perceived voice changes (n=28). Auditory-perceptual assessments and acoustic analysis were performed. Statistical tests were used to correlate the data: ANOVA Test for repeated measures, Friedman Test, Wilcoxon Test, and Pearson's Chi-Square Test. For all tests, the significance level was set at 5%. Results There was a predominance of moderate dysphonia in both groups, according to the auditory-perceptual judgment. There was no statistically significant difference between the groups in the assessment of the auditory-perceptual analysis regarding voice changes (improved, worsened, and unaltered voices) before and after the different technique performance times. Most older women improved their voice after 1 minute of performing the technique. Conclusion Older women often have voice changes when considering the perceptual judgment of the voice. There was no scientific evidence as to the ideal time to obtain a better effect on older women's voices.

Different Performances of Machine Learning Models to Classify Dysphonic and Non-Dysphonic Voices

To analyze the performance of 10 different machine learning (ML) classifiers for discrimination between dysphonic and non-dysphonic voices, using a variance threshold as a method for the selection and reduction of acoustic measurements used in the classifier. We analyzed 435 samples of individuals (337 female and 98 male), with a mean age of 41.07 ± 13.73 years, of which 384 were dysphonic and 51 were non-dysphonic. From the sustained /ε/ vowel sample, 34 acoustic measurements were extracted, including traditional perturbation and noise measurements, cepstral/spectral measurements, and measurements based on nonlinear models. The variance method was used to select the best set of acoustic measurements. We tested the performance of the best-selected set with 10 ML classifiers using precision, sensitivity, specificity, accuracy, and F1-Score measurements. The kappa coefficient was used to verify the reproducibility between the two datasets (training and testing). The naive Bayes (NB) and stochastic gradient descent classifier (SGDC) models performed best in terms of accuracy, AUC, sensitivity, and specificity for a reduced dataset of 15 acoustic measures compared to the full dataset of 34 acoustic measures. SGDC and NB obtained the best performance results, with an accuracy of 0.91 and 0.76, respectively. These two classifiers presented moderate agreement, with a Kappa of 0.57 (SGDC) and 0.45 (NB). Among the tested models, the NB and SGDC models performed better in discriminating between dysphonic and non-dysphonic voices from a set of 15 acoustic measures.

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.

The Mongolian Vowel Acoustic Model Based on the Clustering Algorithm.

In order to solve the problem of vowel acoustic modeling in the Mongolian language and provide more scientific core technical support for speech recognition system, a vowel acoustic model based on the clustering algorithm and speech recognition technology was proposed. The language vowel recognition system and Mongolian speech recognition system are constructed, the Mongolian vowel acoustic database is mainly designed, and the number of vowel samples is constantly increased to ensure the acoustic and linguistic phenomena that may be encountered in the Mongolian vowel database. At the same time, classification modeling and context modeling are used to improve the accuracy of the acoustic model. Through experiments, it is found that the experimental results of sparse tritones show that the model has the highest recognition accuracy of 45% for sentences and 86% for words, which is more than 2% higher than before, providing some technical support for Mongolian learning and pronunciation.

Voice Quality in Telephone Interviews: A preliminary Acoustic Investigation

To investigate the impact of standardized mobile phone recordings passed through a telecom channel on acoustic markers of voice quality and on its perception by voice experts in normophonic speakers. Continuous speech and a sustained vowel were recorded for fourteen female and ten male normophonic speakers. The recordings were done simultaneously with a head-mounted high-quality microphone and through the telephone network on a receiving smartphone. Twenty-two acoustic voice quality, breathiness and pitch-related measures were extracted from the recordings. Nine vocologists perceptually rated the G, R and B parameters of the GRBAS scale on each voice sample. The reproducibility, the recording type, the stimulus type and the gender effects, as well as the correlation between acoustic and perceptual measures were investigated. The sustained vowel samples are damped after one second. Only the frequencies between 100 and 3700Hz are passed through the telecom channel and the frequency response is characterized by peaks and troughs. The acoustic measures show a good reproducibility over the three repetitions. All measures significantly differ between the recording types, except for the local jitter, the harmonics-to-noise ratio by Dejonckere and Lebacq, the period standard deviation and all six pitch measures. The AVQI score is higher in telephone recordings, while the ABI score is lower. Significant differences between genders are also found for most of the measures; while the AVQI is similar in men and women, the ABI is higher in women in both recording types. For the perceptual assessment, the interrater agreement is rather low, while the reproducibility over the three repetitions is good. Few significant differences between recording types are observed, except for lower breathiness ratings on telephone recordings. G ratings are significantly more severe on the sustained vowel on both recording types, R ratings only on telephone recordings. While roughness is rated higher in men on telephone recordings by most experts, no gender effect is observed for breathiness on either recording types. Finally, neither the AVQI nor the ABI yield strong correlations with any of the perceptual parameters. Our results show that passing a voice signal through a telecom channel induces filter and noise effects that limit the use of common acoustic voice quality measures and indexes. The AVQI and ABI are both significantly impacted by the recording type. The most reliable acoustic measures seem to be pitch perturbation (local jitter and period standard deviation) as well as the harmonics-to-noise ratio from Dejonckere and Lebacq. Our results also underline that raters are not equally sensitive to the various factors, including the recording type, the stimulus type and the gender effects. Neither of the three perceptual parameters G, R and B seem to be reliably measurable on telephone recordings using the two investigated acoustic indexes. Future studies investigating the impact of voice quality in telephone conversations should thus focus on acoustic measures on continuous speech samples that are limited to the frequency response of the telecom channel and that are not too sensitive to environmental and additive noise.

Detection of Glottic Neoplasm Based on Voice Signals Using Deep Neural Networks

Use of artificial intelligence in medical practice has considerably increased in recent years. Machine learning algorithms have been successfully applied to detect abnormal voice samples. This letter further investigated the ability of deep neural networks (DNNs) to differentiate voice signals between patients with glottic neoplasms and those with benign voice disorders. We retrospectively included voice samples from 43 patients with histologically proved glottic neoplasm and 129 patients with benign voice disorders. Each neoplastic case was matched with three benign cases (i.e., vocal palsy, atrophy, and phonotraumatic lesions) based on age, sex, and the date of first clinical visit. Mel-frequency cepstrum coefficients were retrieved from the voice samples of a continuous vowel “ah.” The optimal DNN configuration of hidden layers and neurons was systematically investigated. The highest training accuracy (98%) was achieved at four hidden layers, with 500 neurons per layer. By threefold cross validation, the testing accuracy was 86.11% with a detection threshold of 0.2, whereas the sensitivity, specificity, and the area under the receiver operating characteristic curve were 77.78%, 88.89%, and 0.91%, respectively. These results demonstrated that DNNs are a potential tool for detecting and differentiating glottic neoplasms from common benign voice disorders.

The Effect of Microphone Frequency Response on Spectral and Cepstral Measures of Voice: An Examination of Low-Cost Electret Headset Microphones.

The purpose of this study was to establish the frequency response of a selection of low-cost headset microphones that could be given to subjects for remote voice recordings and to examine the effect of microphone type and frequency response on key acoustic measures related to voice quality obtained from speech and vowel samples. The frequency responses of three low-cost headset microphones were evaluated using pink noise generated via a head-and-torso model. Each of the headset microphones was then used to record a series of speech and vowel samples prerecorded from 24 speakers who represented a diversity of sex, age, fundamental frequency (F o), and voice quality types. Recordings were later analyzed for the following measures: smoothed cepstral peak prominence (CPP; dB), low versus high spectral ratio (L/H ratio; dB), CPP F o (Hz), and cepstral spectral index of dysphonia (CSID). The frequency response of the microphones under test was observed to have nonsignificant effects on measures of the CPP and CPP F o, significant effects on the CSID in speech contexts, and strong and significant effects on the measure of spectral tilt (L/H ratio). However, the correlations between the various headset microphones and a reference precision microphone were excellent (rs > .90). The headset microphones under test all showed the capability to track a wide range of diversity in the voice signal. Though the use of higher quality microphones that have demonstrated specifications is recommended for typical research and clinical purposes, low-cost electret microphones may be used to provide valid measures of voice, specifically when the same microphone and signal chain is used for the evaluation of pre- versus posttreatment change or intergroup comparisons.