Statistical Modeling of Speech Spectral Coefficients in Patients with Parkinson's Disease

To automatically detect and monitor Parkinson's disease (PD) from speech, crafting features which robustly differentiate between speech of PD patients and healthy speakers is necessary. Since the voice of PD patients is typically breathy and semi-whispery and since their speech is characterized by significantly less pauses than healthy speech, it can be expected that PD speech spectral coefficients are less super-Gaussian than healthy speech spectral coefficients. In this paper we propose to use the distribution of speech spectral coefficients as a novel discriminatory feature between PD and healthy speech. Speech spectral magnitudes are modeled using theWeibull distribution, with the shape parameter controlling the super-Gaussianity of the complex spectral coefficients. Supported by empirical analysis on healthy and PD speech, it is shown that the shape parameter modeling PD spectral magnitudes is larger than the shape parameter modeling healthy spectral magnitudes, i.e., PD spectral coefficients are less super-Gaussian than healthy spectral coefficients. This result should be taken into account not only when discriminating between healthy and PD speech, but also when developing statistical signal processing techniques.