Determining the Severity of Depression in Speech Based on Combination of Acoustic-Space and Score-Space Features

Mental diseases like depression create disorders in vocal limbs and artificial intelligence researchers believe that they can help psychologists and psychiatrists to measure depression level based on voice features in the patient’s speech. The proposed method in this paper combines acoustic-space and score-space features to estimate Beck’s Depression Index (BDI-II) using several regression techniques. To this end, Mel Frequency Cepstral Coefficient (MFCC) features and glottal waveform, pitch, perceptual linear predictive cepstral coefficients (PLPC), IIR-CQT Frequency Cepstral Coefficients (ICMC), Minimum-Variance Distortionless Response (MVDR), Gabor filter bank features (GBFB), and separated Gabor filter bank features (SGBFB) along with the scores obtained from each and all of them was used individually and in combination. The results indicated that combining Long-Term ICMC and Long-Term MFCC features and their GMM-based scores led to the best estimate. In addition, based on the AVEC2014 open-access database, the approach adopted here outperformed those of many other studies so that a lower mean absolute error (MAE) was obtained.