Non-Linear and Non-Stationary Speech Analysis of Parkinson’s Disease Using Empirical Mode Decomposition

Abstract

<p>The objective of the thesis is to provide a set of features that represents the physiological manifestation of Parkinson's disease (PD) in voice and machine learning methods to determine PD voice. PD is the only neurological disorder with increasing age-specific prevalence between 1990 and 2015. There is no cure for PD. Early detection can slow down disease progress through treatments. PD voice impairment can occur as early as 7-11 years prior to diagnosis. Parkinsonian dysarthria has a set of well-established hand-crafted features. However, a lot of these features require manual processes by skilled personnel. Furthermore, most PD datasets are too small for deep learning models. </p> <p>This thesis proposes Empirical Mode Decomposition (EMD) to extract non-linear and non-stationary characteristics of PD voice. To assist with automatic feature extraction, a Minimum Spline Enveloping technique was proposed to provide better enveloping on extremely dynamic PD speech. An introduction of PD voice characteristics, analyses of PD voice, and discriminative ability of Intrinsic Mode Functions (IMFs) in downgraded toll-quality voice were provided to establish the basis of the study. A basic set of EMD features was proposed to represent the phonatory characteristic of the sustained vowel produced by PD patients. These features were tested on the large unlabelled mPower corpus using clustering as unsupervised learning. A set of EMD dyadic features was proposed to represent the articulatory features and tested on /pa-ta-ka/ utterance from the PC-GITA database.</p> <p>Segmentation strategies were also compared to see the efficacy of the dyadic features on /pa-ta-ka/ and was found that the standard voice-onset-time and onset-time segmentation does not work well using EMD. Comparing fixed frame size and /pa-ta-ka/ triad segmentations, /pa-ta-ka/ triad outperformed fixed frame size. Using /pa-ta-ka/ utterances, the EMD dyadic feature alone was able to achieve 78% accuracy, which is 8% higher than using a combination of hand-crafted and basic EMD features on sustained /a/ and various diadochokinesia (DDK) utterances. Extension studies on EMD using deep neural networks to approximate the EMD filter-bank to parallelize the sifting process and the possibility of using EMD for motor analysis have been investigated.</p>