Analysis of paralinguistic properties of speech for near-term suicidal risk assessment

The respiratory, phonatory and articulatory processes that are involved in speech production are the result of the coordination of the muscular activity in the respiratory organs, the intrinsic and extrinsic laryngeal muscles, and a large number of articulators. The neuromuscular control of these movements must be very finely tuned to ensure the smoothness of the vocal fold vibrations and the dynamic adjustments from one articulatory setting to the next. Nonlinguistic properties of speech have previously been reported to be highly sensitive towards the physiological irregularities caused by psychomotor disturbances during states of psychological arousal, which produces changes in the speech production mechanism by affecting the respiratory, phonatory, and articulatory processes that in turn are encoded in the speech signal.; Vocal cues that represent the disruptions in the smoothness of the coordination and the tenseness of the muscles involved in the speech production process during states of psychological arousal have been identified as unique and defining symptoms of affective disorders. In this dissertation we investigated the validity of using the vocal properties of speech as a diagnostic tool for the assessment of near-term suicidal risk, which can provide a valuable quantitative supplement to clinical judgment. Statistical analyses were performed on both source (excitation) and filter (vocal tract) domain speech features to determine if the acoustical properties of speech change with high-risk, near-term suicidality. Source domain analysis explored the significance of two excitation-based speech features; vocal jitter (a measure of the variations found within successive periods of the laryngeal vibratory pattern) and slope of the glottal flow spectrum (glottal spectral slope). Filter domain analysis explored the significance of first four mel-cepstral filter bank coefficients, which parameterize the spectral envelope shape. Results of the statistical analyses on thirty patients indicate that physiological irregularities caused during high-risk, near-term suicidal states are reflected on the extracted speech features. Maximum Likelihood classification analyses based on the integrated source and excitation domain features yielded 88.3% correct classification performance among three diagnostic classes (i.e., near-term suicidal, major depresses and non-depressed control).