Biomedical Signal Processing For Bioacoustic Features Extraction and Reproducibility Evaluation

Abstract

Human speech produces acoustic waves that carry information about the speaker’s gender, physiological condition, and pathophysiological state. Bio-acoustic properties obtained by speech signal processing show promise for the analysis of psychiatric illnesses. Alterations of acoustic measures associated with major depressive disorder (MDD) could potentially provide objective biomarkers for depression detection. Understanding bio-acoustic features stability is essential to best design sampling and analysis frameworks. Still, the impact of sample duration on bio-acoustic features’ reproducibility has not been systematically explored. Classification performance of depressed and non-depressed bio-acoustic features measured at different speech durations remains to be investigated. This thesis evaluates the reproducibility of bio-acoustic features against changes in speech durations and speech tasks in depressed and non-depressed English speakers. It also investigates the classification potential, in a binary manner, of bio-acoustic features quantified at short speech durations for MDD detection. Thus, source, spectral shape, cepstral, prosodic, and formants features were extracted from speech signals. The intraclass correlation coefficients were calculated to measure feature reproducibility. Support vector machines with radial basis function kernel were employed to evaluate the effect of speech duration on classification performance. Experimental results indicate that the number of reproducible features (out of 125) decreased stepwisely with duration reduction in both depressed and non-depressed speakers. Gender differences had a significant impact on the reproducibility of some features (e.g., pitch). The results also showed a slight improvement in the classification performance (accuracy, weighted F1 score, recall, and precision) when shortening the duration. In conclusion, bio-acoustic characteristics are less reproducible in shorter speech samples and are affected by gender. Classification metrics are also influenced by speech data duration. Designing speech samples and building classification models to potentially assist medical practitioners in depression diagnosis have to consider the duration effects and gender differences.