An investigation of acoustic, linguistic, and affect based methods for speech depression assessment

Given that there is no definitive, widely accepted laboratory test for depression, there is a need for effective new methods to help with patient diagnosis and monitoring. While abnormal speech behaviours are identified as useful indicators for recognizing depression, the effects of depression on speech production warrant further investigation. Automatic approaches for speech-based depression assessment show promise, however they are still in their early developmental phase, and many questions remain unanswered. This thesis investigates which acoustic, linguistic, and affective aspects of speech are discriminative of depression, towards informing speech elicitation methods and providing more effective knowledge-driven automatic speech-based depression disorder classification system designs. Through experimentation herein, it is discovered that as articulatory-gestural effort, linguistic complexity, and affect dimensions increase, more discriminative speech-language differences are exhibited between depressed and non-depressed speakers. By examining vowel articulatory parameters, statistically significant differences in acoustic characteristics are found at a paraphonetic level in depressed speakers, such as shorter vowel durations and less variance for ‘low’, ‘back’, and ‘rounded’ vowel positions. Furthermore, when compared with non-depressed speakers, experiments reveal that depressed speakers exhibit reduced linguistic stress attributes. To help guide speech data selection and exploit articulatory differences between depressed/non- depressed speakers, two new measures based on articulatory acquisition mastery and phoneme transition gestural effort are devised. From a linguistic perspective, experiments herein show that statistically significantly greater numbers of speech disfluencies occur in depressed speakers than non-depressed speakers, especially in terms of hesitations and malapropisms. It is discovered that linguistic placement of affect token words at the beginning of elicited read sentences rather than the middle/end of sentences provides better acoustic discriminative cues for identifying depressed speakers. From an affective perspective, experiments conducted show that individually computing features on a per-valence affect category (negative, neutral, positive) basis and then fusing them into a multi-affect feature results in significant improvements, up to 100% accuracy, in depression classification relative to separate per-affect-range systems or affect-agnostic approaches