Depression Detection Using Automatic Transcriptions of De-Identified Speech

Depression is a mood disorder that is usually addressed by outpatient treatments in order to favour patient’s inclusion in society. This leads to a need for novel automatic tools exploiting speech processing approaches that can help to monitor the emotional state of patients via telephone or the Internet. However, the transmission, processing and subsequent storage of such sensitive data raises several privacy concerns. Speech deidentification can be used to protect the patients’ identity. Nevertheless, these techniques modify the speech signal, eventually affecting the performance of depression detection approaches based on either speech characteristics or automatic transcriptions. This paper presents a study on the influence of speech de-identification when using transcription-based approaches for depression detection. To this effect, a system based on the global vectors method for natural language processing is proposed. In contrast to previous works, two main sources of nuisance have been considered: the de-identification process itself and the transcription errors introduced by the automatic recognition of the patients’ speech. Experimental validation on the DAIC-WOZ corpus reveals very promising results, obtaining only a slight performance degradation with respect to the use of manual transcriptions.