Evaluation of Acoustic Analyses of Voice in Nonoptimized Conditions

Conducting high-quality and reliable acoustic analysis: A tutorial focused on training research assistants

Open science practices have led to an increase in available speech datasets for researchers interested in acoustic analysis. Accurate evaluation of these databases frequently requires manual or semi-automated analysis. The time-intensive nature of these analyses makes them ideally suited for research assistants in laboratories focused on speech and voice production. However, the completion of high-quality, consistent, and reliable analyses requires clear rules and guidelines for all research assistants to follow. This tutorial will provide information on training and mentoring research assistants to complete these analyses, covering areas including RA training, ongoing data analysis monitoring, and documentation needed for reliable and re-creatable findings.

Machine Learning-Assisted Speech Analysis for Early Detection of Parkinson's Disease: A Study on Speaker Diarization and Classification Techniques

Parkinson's disease (PD) is a neurodegenerative disorder characterized by a range of motor and non-motor symptoms. One of the notable non-motor symptoms of PD is the presence of vocal disorders, attributed to the underlying pathophysiological changes in the neural control of the laryngeal and vocal tract musculature. From this perspective, the integration of machine learning (ML) techniques in the analysis of speech signals has significantly contributed to the detection and diagnosis of PD. Particularly, MEL Frequency Cepstral Coefficients (MFCCs) and Gammatone Frequency Cepstral Coefficients (GTCCs) are both feature extraction techniques commonly used in the field of speech and audio signal processing that could exhibit great potential for vocal disorder identification. This study presents a novel approach to the early detection of PD through ML applied to speech analysis, leveraging both MFCCs and GTCCs. The recordings contained in the Mobile Device Voice Recordings at King's College London (MDVR-KCL) dataset were used. These recordings were collected from healthy individuals and PD patients while they read a passage and during a spontaneous conversation on the phone. Particularly, the speech data regarding the spontaneous dialogue task were processed through speaker diarization, a technique that partitions an audio stream into homogeneous segments according to speaker identity. The ML applied to MFCCS and GTCCs allowed us to classify PD patients with a test accuracy of 92.3%. This research further demonstrates the potential to employ mobile phones as a non-invasive, cost-effective tool for the early detection of PD, significantly improving patient prognosis and quality of life.

Observations and Considerations for Implementing Remote Acoustic Voice Recording and Analysis in Clinical Practice

OBJECTIVES/HYPOTHESIS

Remote voice recording and acoustic analysis allow for comprehensive voice assessment and outcome tracking without the requirements of travel to the clinic, in-person visit, or expensive equipment. This paper delineates the process and considerations for implementing remote voice recording and acoustic analysis in a high-volume university voice clinic.

STUDY DESIGN

Clinical Focus.

METHODS

Acoustic voice recordings were attempted on 108 unique patients over a 6-month period using a remote voice recording phone application. Development of the clinical process including determining normative data in which to compare acoustic results, clinician training, and clinical application is described. The treating Speech Language Pathologists (SLPs) were surveyed 2 months after implementation to assess ease of application, identify challenges and assess implementation of potential solutions.

RESULTS

Of 108 unique patients, 83 patients were successful in completing the process of synchronous remote acoustic voice recording in conjunction with their SLP clinician. The process of downloading the application, setting up, and obtaining voice recordings was most commonly 10-20 minutes according to the 8 SLPs surveyed. Challenges and helpful techniques were identified.

CONCLUSIONS

Remote acoustic voice recordings have allowed SLPs to continue to complete a comprehensive voice evaluation in a telepractice model. Given emerging knowledge about the viability of remote voice recordings, the success in obtaining acoustic data remotely, and the accessibility of a low-cost app for SLPs makes remote voice recordings a viable option to facilitate remote clinical care and research investigation.

Characterization of Primary Muscle Tension Dysphonia Using Acoustic and Aerodynamic Voice Metrics

OBJECTIVES/HYPOTHESIS

The objectives of this study were to (1) identify optimal clusters of 15 standard acoustic and aerodynamic voice metrics recommended by the American Speech-Language-Hearing Association (ASHA) to improve characterization of patients with primary muscle tension dysphonia (pMTD) and (2) identify combinations of these 15 metrics that could differentiate pMTD from other types of voice disorders.

STUDY DESIGN

Retrospective multiparametric METHODS: Random forest modeling, independent t-tests, logistic regression, and affinity propagation clustering were implemented on a retrospective dataset of 15 acoustic and aerodynamic metrics.

RESULTS

Ten percent of patients seen at the New York University (NYU) Voice Center over two years met the study criteria for pMTD (92 out of 983 patients), with 65 patients with pMTD and 701 of non-pMTD patients with complete data across all 15 acoustic and aerodynamic voice metrics. PCA plots and affinity propagation clustering demonstrated substantial overlap between the two groups on these parameters. The highest ranked parameters by level of importance with random forest models-(1) mean airflow during voicing (L/sec), (2) mean SPL during voicing (dB), (3) mean peak air pressure (cmH2O), (4) highest F0 (Hz), and (5) CPP mean vowel (dB)-accounted for only 65% of variance. T-tests showed three of these parameters-(1) CPP mean vowel (dB), (2) highest F0 (Hz), and (3) mean peak air pressure (cmH2O)-were statistically significant; however, the log2-fold change for each parameter was minimal.

CONCLUSION

Computational models and multivariate statistical testing on 15 acoustic and aerodynamic voice metrics were unable to adequately characterize pMTD and determine differences between the two groups (pMTD and non-pMTD). Further validation of these metrics is needed with voice elicitation tasks that target physiological challenges to the vocal system from baseline vocal acoustic and aerodynamic ouput. Future work should also place greater focus on validating metrics of physiological correlates (eg, neuromuscular processes, laryngeal-respiratory kinematics) across the vocal subsystems over traditional vocal output measures (eg, acoustics, aerodynamics) for patients with pMTD.

LEVEL OF EVIDENCE

II.

Addressing disparities in speech-language pathology and laryngology services with telehealth

The COVID-19 pandemic disproportionately affected the health and well-being of marginalized communities, and it brought greater awareness to disparities in health care access and utilization. Addressing these disparities is difficult because of their multidimensional nature. Predisposing factors (demographic information, social structure, and beliefs), enabling factors (family and community) and illness levels (perceived and evaluated illness) are thought to jointly contribute to such disparities. Research has demonstrated that disparities in access and utilization of speech-language pathology and laryngology services are the result of racial and ethnic differences, geographic factors, sex, gender, educational background, income level and insurance status. For example, persons from diverse racial and ethnic backgrounds have been found to be less likely to attend or adhere to voice rehabilitation, and they are more likely to delay health care due to language barriers, longer wait times, a lack of transportation and difficulties contacting their physician. The purpose of this paper is to summarize existing research on telehealth, discuss how telehealth offers the potential to eliminate some disparities in the access and utilization of voice care, review its limitations, and encourage continued research in this area. A clinical perspective from a large volume laryngology clinic in a major city in northeastern United States highlights the use of telehealth in the provision of voice care by a laryngologist and speech-language pathologist during and after the COVID19 pandemic.

A Machine-Learning Algorithm for the Automated Perceptual Evaluation of Dysphonia Severity

OBJECTIVES

Auditory-perceptual assessments are the gold standard for assessing voice quality. This project aims to develop a machine-learning model for measuring perceptual dysphonia severity of audio samples consistent with assessments by expert raters.

METHODS

The Perceptual Voice Qualities Database samples were used, including sustained vowel and Consensus Auditory-Perceptual Evaluation of Voice sentences, which were previously expertly rated on a 0-100 scale. The OpenSMILE (audEERING GmbH, Gilching, Germany) toolkit was used to extract acoustic (Mel-Frequency Cepstral Coefficient-based, n = 1428) and prosodic (n = 152) features, pitch onsets, and recording duration. We utilized a support vector machine and these features (n = 1582) for automated assessment of dysphonia severity. Recordings were separated into vowels (V) and sentences (S) and features were extracted separately from each. Final voice quality predictions were made by combining the features extracted from the individual components with the whole audio (WA) sample (three file sets: S, V, WA).

RESULTS

This algorithm has a high correlation (r = 0.847) with estimates of expert raters. The root mean square error was 13.36. Increasing signal complexity resulted in better estimation of dysphonia, whereby combining the features outperformed WA, S, and V sets individually.

CONCLUSION

A novel machine-learning algorithm was able to perform perceptual estimates of dysphonia severity using standardized audio samples on a 100-point scale. This was highly correlated to expert raters. This suggests that ML algorithms could offer an objective method for evaluating voice samples for dysphonia severity.

LEVEL OF EVIDENCE: 4

Different in different ways: A network-analysis approach to voice and prosody in Autism Spectrum Disorder

The current study investigated whether the difficulty in finding group differences in prosody between speakers with autism spectrum disorder (ASD) and neurotypical (NT) speakers might be explained by identifying different acoustic profiles of speakers which, while still perceived as atypical, might be characterized by different acoustic qualities. We modelled the speech from a selection of speakers (N = 26), with and without ASD, as a network of nodes defined by acoustic features. We used a community-detection algorithm to identify clusters of speakers who were acoustically similar and compared these clusters with atypicality ratings by naïve and expert human raters. Results identified three clusters: one primarily composed of speakers with ASD, one of mostly NT speakers, and one comprised of an even mixture of ASD and NT speakers. The human raters were highly reliable at distinguishing speakers with and without ASD, regardless of which cluster the speaker was in. These results suggest that community-detection methods using a network approach may complement commonly-employed human ratings to improve our understanding of the intonation profiles in ASD.

An iOS-based VoiceScreen application: feasibility for use in clinical settings-a pilot study

OBJECTIVES

To elaborate the application suitable for smartphones for estimation of Acoustic Voice Quality Index (AVQI) and evaluate its usability in the clinical setting.

METHODS

An elaborated AVQI automatization and background noise monitoring functions were implemented into a mobile "VoiceScreen" application running the iOS operating system. A study group consisted of 103 adult individuals with normal voices (n = 30) and 73 patients with pathological voices. Voice recordings were performed in the clinical setting with "VoiceScreen" app using iPhone 8 microphones. Voices of 30 patients were recorded before and 1 month after phonosurgical intervention. To evaluate the diagnostic accuracy differentiating normal and pathological voice, the receiver-operating characteristic statistics, i.e., area under the curve (AUC), sensitivity and specificity, and correct classification rate (CCR) were used.

RESULTS

A high level of precision of AVQI in discriminating between normal and dysphonic voices was yielded with corresponding AUC = 0.937. The AVQI cutoff score of 3.4 demonstrated a sensitivity of 86.3% and specificity of 95.6% with a CCR of 89.2%. The preoperative mean value of the AVQI [6.01(SD 2.39)] in the post-phonosurgical follow-up group decreased to 2.00 (SD 1.08). No statistically significant differences (p = 0.216) between AVQI measurements in a normal voice and 1-month follow-up after phonosurgery groups were revealed.

CONCLUSIONS

The "VoiceScreen" app represents an accurate and robust tool for voice quality measurement and demonstrates the potential to be used in clinical settings as a sensitive measure of voice changes across phonosurgical treatment outcomes.

Comparison of Acoustic Voice Features Derived From Mobile Devices and Studio Microphone Recordings

OBJECTIVES/HYPOTHESIS

Improvements in mobile device technology offer new opportunities for remote monitoring of voice for home and clinical assessment. However, there is a need to establish equivalence between features derived from signals recorded from mobile devices and gold standard microphone-preamplifiers. In this study acoustic voice features from android smartphone, tablet, and microphone-preamplifier recordings were compared.

METHODS

Data were recorded from 37 volunteers (20 female) with no history of speech disorder and six volunteers with Huntington's disease (HD) during sustained vowel (SV) phonation, reading passage (RP), and five syllable repetition (SR) tasks. The following features were estimated: fundamental frequency median and standard deviation (F0 and SD F0), harmonics-to-noise ratio (HNR), local jitter, relative average perturbation of jitter (RAP), five-point period perturbation quotient (PPQ5), difference of differences of amplitude and periods (DDA and DDP), shimmer, and amplitude perturbation quotients (APQ3, APQ5, and APQ11).

RESULTS

Bland-Altman analysis revealed good agreement between microphone and mobile devices for fundamental frequency, jitter, RAP, PPQ5, and DDP during all tasks and a bias for HNR, shimmer and its variants (APQ3, APQ5, APQ11, and DDA). Significant differences were observed between devices for HNR, shimmer, and its variants for all tasks. High correlation was observed between devices for all features, except SD F0 for RP. Similar results were observed in the HD group for SV and SR task. Biological sex had a significant effect on F0 and HNR during all tests, and for jitter, RAP, PPQ5, DDP, and shimmer for RP and SR. No significant effect of age was observed.

CONCLUSIONS

Mobile devices provided good agreement with state of the art, high-quality microphones during structured speech tasks for features derived from frequency components of the audio recordings. Caution should be taken when estimating HNR, shimmer and its variants from recordings made with mobile devices.

Developing an Artificial Intelligence Tool to Predict Vocal Cord Pathology in Primary Care Settings

OBJECTIVES

Diagnostic tools for voice disorders are lacking for primary care physicians. Artificial intelligence (AI) tools may add to the armamentarium for physicians, decreasing the time to diagnosis and limiting the burden of dysphonia.

METHODS

Voice recordings of patients were collected from 2019 to 2021 using smartphones. The Saarbruecken dataset was included for comparison. Audio files were converted to mel-spectrograms using TensorFlow. Diagnostic categories were created to group pathology, including neurological and muscular disorders, inflammatory, mass lesions, and normal. The samples were further separated into sustained/a/and the rainbow passage.

RESULTS

Two hundred three prospective samples and 1131 samples were used from the Saarbruecken database. The AI detected abnormal pathology with an F1-score of 98%. The artificial neural network (ANN) differentiated key pathologies, including unilateral paralysis, laryngitis, adductor spasmodic dysphonia (ADSD), mass lesions, and normal samples with 39%-87% F-1 scores. The Calgary database models had higher F-1 scores in a head-to-head comparison to the Saarbruecken and combined datasets (87% vs. 58% and 50%). The AI outperformed otolaryngologists using a standardized test set of recordings (83% compared to 55% ± 15%).

CONCLUSION

An AI tool was created to differentiate pathology by individual or categorical diagnosis with high evaluation metrics. Prospective data should be collected in a controlled fashion to reduce intrinsic variability between recordings. Multi-center data collaborations are imperative to increase the prediction capability of AI tools for detecting vocal cord pathology. We provide proof-of-concept for an AI tool to assist primary care physicians in managing dysphonic patients.

LEVEL OF EVIDENCE

3 Laryngoscope, 2022.

A Practical Guide to Calculating Cepstral Peak Prominence in Praat

The acoustic measure of cepstral peak prominence (CPP) is recommended for the analysis of dysphonia. Yet, clinical use of this measure is not universal, as clinicians and researchers are still learning the strengths and limitations of this measure. Furthermore, affordable access to specialized acoustic software is a significant barrier to universal CPP use. This article will provide a guide on how to calculate CPP in Praat, a free software program, using a new CPP plugin. Important external factors that could influence CPP measures are discussed, and suggestions for clinical use are provided. As CPP becomes more widely used by clinicians and researchers, it is important to consider external factors that may inadvertently influence CPP values. Controlling for these external factors will aid in reducing variability across CPP values, which will make CPP a valuable tool for both clinical and research purposes.

Acoustic Variability in the Healthy Female Voice Within and Across Days: How Much and Why?

Purpose The aims of this study were (1) to quantify variability in voice production (as measured acoustically) within and across consecutive days in vocally healthy female speakers, (2) to identify which acoustic measures are sensitive to this variability, and (3) to identify participant characteristics related to such voice variability. Method Participants included 45 young women with normal voices who were stratified by age, specifically 18-23, 24-29, and 30-35 years. Following an initial acoustic and auditory-perceptual voice assessment, participants performed standardized field voice recordings 3 times daily across a 7-day period. Acoustic analyses involved 32 cepstral-, spectral-, and time-based measures of connected speech and sustained vowels. Relationships among acoustic data and select demographic, health, and lifestyle (i.e., participant-based) factors were also examined. Results Significant time-of-day effects were observed for acoustic analyses within speakers (p < .05), with voices generally being worse in the morning. No significant differences were observed across consecutive days. Variations in voice production were associated with several participant factors, including improved voice with increased voice use; self-perceived poor voice function, minimal or no alcohol consumption, and extroverted personality; and worse voice with regular or current menstruation, depression, and anxiety. Conclusions This acoustic study provides essential information regarding the nature and extent to which healthy voices vary throughout the day and week. Participant-based factors that were associated with improved voice over time included increased voice use, self-perceived poor voice function, minimal or no alcohol consumption, and extroverted personality. Factors associated with worse voice production over time included regular or current menstruation, and depression and anxiety.