Differential performance of automatic speech-based depression classification across smartphones

摘要

Smartphones offer exciting new prospects for measuring and monitoring mental health. While research into speech-based analysis of depression shows promise for smartphone app deployment, systematic studies investigating how smartphone device variability affects performance of speech-based classification of health conditions like depression have yet to be reported. Differing audio acquisition techniques between different devices introduces variability into the speech-based depression classification problem. Experiments reported herein reveal dissimilarities in depression classification performance among Android™ smartphones, particularly for spectral features. This preliminary study on speech-based depression classification shows that by using smartphone version-specific models, relatively channel-independent features, and/or normalization methods unwanted performance variability can be mitigated - producing significant improvement over a channel-agnostic feature approach.