End-to-end deep learning models have shown promising results for the automatic screening of Parkinson’s disease by voice and speech. However, these models often suffer degradation in their performance when applied to scenarios involving multiple corpora. In addition, they also show corpus-dependent clusterings. These facts indicate a lack of generalisation or the presence of certain shortcuts in the decision, and also suggest the need for developing new corpus-independent models. In this respect, this work explores the use of domain adversarial training as a viable strategy to develop models that retain their discriminative capacity to detect Parkinson’s disease across diverse datasets. The paper presents three deep learning architectures and their domain adversarial counterparts. The models were evaluated with sustained vowels and diadochokinetic recordings extracted from four corpora with different demographics, dialects or languages, and recording conditions. The results showed that the space distribution of the embedding features extracted by the domain adversarial networks exhibits a higher intra-class cohesion. This behaviour is supported by a decrease in the variability and inter-domain divergence computed within each class. The findings suggest that domain adversarial networks are able to learn the common characteristics present in Parkinsonian voice and speech, which are supposed to be corpus, and consequently, language independent. Overall, this effort provides evidence that domain adaptation techniques refine the existing end-to-end deep learning approaches for Parkinson’s disease detection from voice and speech, achieving more generalizable models.

The ambulatory assessment of vocal function can be significantly enhanced by having access to physiologically based features that describe underlying pathophysiological mechanisms in individuals with voice disorders. This type of enhancement can improve methods for the prevention, diagnosis, and treatment of behaviorally based voice disorders. Unfortunately, the direct measurement of important vocal features such as subglottal pressure, vocal fold collision pressure, and laryngeal muscle activation is impractical in laboratory and ambulatory settings. In this study, we introduce a method to estimate these features during phonation from a neck-surface vibration signal through a framework that integrates a physiologically relevant model of voice production and machine learning tools. The signal from a neck-surface accelerometer is first processed using subglottal impedance-based inverse filtering to yield an estimate of the unsteady glottal airflow. Seven aerodynamic and acoustic features are extracted from the neck surface accelerometer and an optional microphone signal. A neural network architecture is selected to provide a mapping between the seven input features and subglottal pressure, vocal fold collision pressure, and cricothyroid and thyroarytenoid muscle activation. This non-linear mapping is trained solely with 13,000 Monte Carlo simulations of a voice production model that utilizes a symmetric triangular body-cover model of the vocal folds. The performance of the method was compared against laboratory data from synchronous recordings of oral airflow, intraoral pressure, microphone, and neck-surface vibration in 79 vocally healthy female participants uttering consecutive /pæ/ syllable strings at comfortable, loud, and soft levels. The mean absolute error and root-mean-square error for estimating the mean subglottal pressure were 191 Pa (1.95 cm H₂O) and 243 Pa (2.48 cm H₂O), respectively, which are comparable with previous studies but with the key advantage of not requiring subject-specific training and yielding more output measures. The validation of vocal fold collision pressure and laryngeal muscle activation was performed with synthetic values as reference. These initial results provide valuable insight for further vocal fold model refinement and constitute a proof of concept that the proposed machine learning method is a feasible option for providing physiologically relevant measures for laboratory and ambulatory assessment of vocal function.

The purpose of this paper is to report on the first in vivo application of a recently developed transoral, dual-sensor pressure probe that directly measures intraglottal, subglottal, and vocal fold collision pressures during phonation. Synchronous measurement of intraglottal and subglottal pressures was accomplished using two miniature pressure sensors mounted on the end of the probe and inserted transorally in a 78-year-old male who had previously undergone surgical removal of his right vocal fold for treatment of laryngeal cancer. The endoscopist used one hand to position the custom probe against the surgically medialized scar band that replaced the right vocal fold and used the other hand to position a transoral endoscope to record laryngeal high-speed videoendoscopy of the vibrating left vocal fold contacting the pressure probe. Visualization of the larynx during sustained phonation allowed the endoscopist to place the dual-sensor pressure probe such that the proximal sensor was positioned intraglottally and the distal sensor subglottally. The proximal pressure sensor was verified to be in the strike zone of vocal fold collision during phonation when the intraglottal pressure signal exhibited three characteristics: an impulsive peak at the start of the closed phase, a rounded peak during the open phase, and a minimum value around zero immediately preceding the impulsive peak of the subsequent phonatory cycle. Numerical voice production modeling was applied to validate model-based predictions of vocal fold collision pressure using kinematic vocal fold measures. The results successfully demonstrated feasibility of in vivo measurement of vocal fold collision pressure in an individual with a hemilaryngectomy, motivating ongoing data collection that is designed to aid in the development of vocal dose measures that incorporate vocal fold impact collision and stresses.

The aerodynamic voice assessment of subglottal air pressure can discriminate between speakers with typical voices from patients with voice disorders, with further evidence validating subglottal pressure as a clinical outcome measure. Although estimating subglottal pressure during phonation is an important component of a standard voice assessment, current methods for estimating subglottal pressure rely on non-natural speech tasks in a clinical or laboratory setting. This study reports on the validation of a method for subglottal pressure estimation in individuals with and without voice disorders that can be translated to connected speech to enable the monitoring of vocal function and behavior in real-world settings. During a laboratory calibration session, a participant-specific multiple regression model was derived to estimate subglottal pressure from a neck-surface vibration signal that can be recorded during natural speech production. The model was derived for vocally typical individuals and patients diagnosed with phonotraumatic vocal fold lesions, primary muscle tension dysphonia, and unilateral vocal fold paralysis. Estimates of subglottal pressure using the developed method exhibited significantly lower error than alternative methods in the literature, with average errors ranging from 1.13 to 2.08 cm H₂O for the participant groups. The model was then applied during activities of daily living, thus yielding ambulatory estimates of subglottal pressure for the first time in these populations. Results point to the feasibility and potential of real-time monitoring of subglottal pressure during an individual’s daily life for the prevention, assessment, and treatment of voice disorders.

Many voice disorders are linked to imbalanced muscle activity and known to exhibit asymmetric vocal fold vibration. However, the relation between imbalanced muscle activation and asymmetric vocal fold vibration is not well understood. This study introduces an asymmetric triangular body-cover model of the vocal folds, controlled by the activation of bilateral intrinsic laryngeal muscles, to investigate the effects of muscle imbalance on vocal fold oscillation. Various scenarios were considered, encompassing imbalance in individual muscles and muscle pairs, as well as accounting for asymmetry in lumped element parameters. Measurements of amplitude and phase asymmetries were employed to match the oscillatory behavior of two pathological cases: unilateral paralysis and muscle tension dysphonia. The resulting simulations exhibit muscle imbalance consistent with expectations in the composition of these voice disorders, yielding asymmetries exceeding 30% for paralysis and below 5% for dysphonia. This underscores the relevance of muscle imbalance in representing phonatory scenarios and its potential for characterizing asymmetry in vocal fold vibration.

This study presents a novel application of a Probabilistic Bayesian Neural Network (PBNN) for estimating vocal function variables and enhancing non-invasive ambulatory voice monitoring by addressing aleatoric and epistemic uncertainties in regression tasks. The proposed PBNN allows for estimating key physiological parameters including subglottal pressure, vocal fold contact pressure, thyroarytenoid, and cricothyroid muscle activations, from seven aerodynamic and acoustic features. The PBNN is trained on the Triangular Body-Cover Model (TBCM) of the vocal folds to produce a non-linear inverse mapping between its inputs and outputs. Furthermore, the selected aerodynamic and acoustic features can be obtained in ambulatory settings, thus enhancing the practical applicability of the proposed method. Transfer Learning is then applied to integrate real voice data into the initially synthetic-trained network to refine subglottal pressure estimations. The confidence intervals generated by the PBNN illustrate its ability to identify uncertain estimations, as the results show correlations between prediction errors and the estimated aleatoric and epistemic uncertainties. This correlation is advantageous because it shows that the network can effectively predict potential inaccuracies in its estimations. Increased uncertainty is mainly observed at operating points where the TBCM is likely to exhibit non-linear behaviors, at higher subglottal pressures. This suggests that the selected input features may not be robust enough for capturing the nonlinear effects in the TBCM. These results highlight the potential for future research to assess the viability of incorporating new features and additional measurements that could better capture non-linear responses.