Doctor of Philosophy
Electrical and Computer Engineering
Parkinson’s disease (PD) is the second most common neurodegenerative disease. Statistics show that nearly 90% of people impaired with PD develop voice and speech disorders. Speech production impairments in PD subjects typically result in hypophonia and consequently, poor speech signal-to-noise ratio (SNR) in noisy environments and inferior speech intelligibility and quality. Assessment, monitoring, and improvement of the perceived quality and intelligibility of Parkinsonian voice and speech are, therefore, paramount. In the first study of this thesis, the perceived quality of sustained vowels produced by PD patients was assessed through objective predictors. Subjective quality ratings of sustained vowels were collected from 51 PD patients, on and off the Levodopa medication, and 7 control subjects. Features extracted from the sustained vowel recordings were combined using linear regression (LR) and support vector regression (SVR). An objective metric that combined linear prediction and harmonicity features resulted in a high correlation of 0.81 with subjective ratings, higher than the performance reported in the literature. The second study focused on the prediction of amplified Parkinsonian speech quality. Speech amplifiers are used by PD patients to counteract hyperphonia. To benchmark the amplifier performance, subjective ratings of the quality of speech samples from 11 PD patients and 10 control subjects using 7 different speech amplifiers in different background noise conditions were collected. Objective quality predictors were then developed in combination with machine learning algorithms such as deep learning (DL). It was shown that the speech amplifiers differentially affect Parkinsonian speech quality and that the composite objective metric resulted in a correlation of 0.85 with subjective speech quality ratings. In the third study, a new signal-to-noise feedback (SNF) device was designed and developed to help PD patients control their speech SNR, intelligibility, and quality. The proposed SNF device contained dual ear-level microphones for estimating the speech SNR, a throat accelerometer for reliable voice activity detection, and visual/auditory alarms when the produced speech was below a certain threshold. Performance evaluation of this device in noisy environments demonstrated significant improvements in speech SNR, perceived intelligibility, and predicted quality, especially in high background noise levels.
Summary for Lay Audience
Nearly 90% of people impaired with Parkinson’s disease (PD) develop voice and speech disorders during the course of their disease. Parkinsonian speech is typically accompanied with deterioration in loudness, intelligibility (i.e. how many words of the sentence can the listener understand?), and quality (a multi-dimensional perceptual phenomenon that encompasses attributes such as clarity, pleasantness, and naturalness). Therefore, there is a clinical need to assess the speech quality for people impaired with PD. Traditionally, voice and speech quality are evaluated by a panel of listeners (subjective assessment). While this may be the gold standard, objective estimation of speech quality through computational models is more robust and time and cost-efficient approach. This thesis focuses on the development and evaluation of such objective Parkinsonian speech and voice quality estimators. In this study's first contribution, multiple objective metrics are developed to assess the quality of the Parkinsonian sustained vowels. First, acoustic features of the vowels’ records are extracted to estimate the quality. There is a need to machine learning algorithms to map the acoustic characteristics to the behavioral assessment of speech quality. Investigations in this study led to an automatic quality estimator that predicts the Parkinsonian voice quality with a correlation value (a mathematical measure of similarity) of 0.81 with the subjective scores. In the second contribution of this study, another set of objective metrics are developed to assess the quality of Parkinsonian running speech. This research study also benchmarked the effect of speech amplifiers on Parkinsonian speech quality. Different acoustic features and various and more sophisticated machine learning algorithms like deep learning are used to extract these objective metrics. The correlation value between the subjective and objective estimations of the Parkinsonian speech quality reached 0.85. In the third contribution of this study, a new device is developed and presented to help people impaired with PD to enhance and improve their speech quality and intelligibility. This device is designed to be portable, cost-effective, and easy to build. The performance of the device has been tested in noisy environments. The results showed enhancements of the intelligibility and quality of subjects' speech with the help of the device.
Gaballah, Amr, "Parkinsonian Speech and Voice Quality: Assessment and Improvement" (2019). Electronic Thesis and Dissertation Repository. 6433.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Available for download on Saturday, August 22, 2020