Voice Feature Selection to Improve Performance of Machine Learning Models for Voice Production Inversion

Deep Learning for Neuromuscular Control of Vocal Source for Voice Production

A computational neuromuscular control system that generates lung pressure and three intrinsic laryngeal muscle activations (cricothyroid, thyroarytenoid, and lateral cricoarytenoid) to control the vocal source was developed. In the current study, LeTalker, a biophysical computational model of the vocal system was used as the physical plant. In the LeTalker, a three-mass vocal fold model was used to simulate self-sustained vocal fold oscillation. A constant/ǝ/vowel was used for the vocal tract shape. The trachea was modeled after MRI measurements. The neuromuscular control system generates control parameters to achieve four acoustic targets (fundamental frequency, sound pressure level, normalized spectral centroid, and signal-to-noise ratio) and four somatosensory targets (vocal fold length, and longitudinal fiber stress in the three vocal fold layers). The deep-learning-based control system comprises one acoustic feedforward controller and two feedback (acoustic and somatosensory) controllers. Fifty thousand steady speech signals were generated using the LeTalker for training the control system. The results demonstrated that the control system was able to generate the lung pressure and the three muscle activations such that the four acoustic and four somatosensory targets were reached with high accuracy. After training, the motor command corrections from the feedback controllers were minimal compared to the feedforward controller except for thyroarytenoid muscle activation.

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

OBJECTIVE

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.

METHOD

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).

RESULTS

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).

CONCLUSION

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.

An Improved Glottal Flow Model Based on Seq2Seq LSTM for Simulation of Vocal Fold Vibration

OBJECTIVES

An improved data-driven glottal flow model for fluid-structure interaction (FSI) simulation of the vocal fold vibration is proposed in this paper. This model aims to improve the prediction performance of the previously developed deep neural network (DNN) based empirical flow model (EFM)1 on accuracy and efficiency.

METHODS

A Seq2Seq long short-term memory (LSTM) network is employed in the present model to infer the flow rate and pressure distribution from the subglottal pressure and cross-section area distribution of the glottis. The training data is collected from the generalized glottal shape library generated in Zhang et al.1 RESULTS AND CONCLUSIONS: Compared to the EFM, the present model not only discards the time-consuming optimization process, but also drastically reduces the errors, therefore the prediction performance can be greatly improved. The present model is evaluated by coupling with a solid dynamics solver for FSI simulation, and the results demonstrate a great improvement on accuracy and efficiency.

Estimating subglottal pressure and vocal fold adduction from the produced voice in a single-subject study (L)

We previously reported a simulation-based neural network for estimating vocal fold properties and subglottal pressure from the produced voice. This study aims to validate this neural network in a single-human subject study. The results showed reasonable accuracy of the neural network in estimating the subglottal pressure in this particular human subject. The neural network was also able to qualitatively differentiate soft and loud speech conditions regarding differences in the subglottal pressure and degree of vocal fold adduction. This simulation-based neural network has potential applications in identifying unhealthy vocal behavior and monitoring progress of voice therapy or vocal training.