Preserving Mental Health Information in Speech Anonymization

Widespread speech technologies may change the dynamics of mental health (MH) diagnosis, monitoring, and care by tracking spectrotemporal patterns of speech in a personalized manner. Yet, these technologies have given rise to a large public debate about the ability of users to protect their privacy. This work examines a speech anonymization algorithm that preserves information that is diagnostic of the MH condition of the speaker and representative of the phonological content of speech, while suppressing characteristics related to the speaker identity. The proposed anonymization algorithm relies on an auto-encoder that conducts an identity mapping between the original input, represented as the superposition of Mel-spectrogram coefficients and posteriogram (PPG) vectors, and the reconstructed output. The auto-encoder is trained in an adversarial manner to minimize the loss corresponding to MH-related information and maximize the loss corresponding to the identity (ID) of a speaker. The proposed algorithm is evaluated on data from healthy participants and patients with depression from the Distress Analysis Interview Corpus Wizard of Oz (DAIC-WoZ) dataset. Results indicate that the speech signals synthesized by the proposed anonymization algorithm have higher word error rate (WER), as calculated between the original speech transcripts and transcripts from the anonymized speech signals obtained via an automatic speech recognizer (ASR), and are rated as being less comprehensible by MTurk listeners compared to the original signals. However, the intelligibility of the synthesized speech appears equivalent to baseline speech anonymization algorithms that solely suppress the speaker ID without considering MH-related information, or rely on a cascade of signal processing methods that perform consecutive speech transformations. The proposed algorithm can further effectively suppress information related to the speaker ID, since the corresponding anonymized signals achieve 15.9% reduction in speaker classification accuracy compared to the original speech signals. Finally, the anonymized speech signals yield similar performance to the original speech when they are used as an input for estimating degree of depression severity, and superior performance compared to conventional speech anonymization algorithms that do not consider the preservation of MH-related information as an optimization criterion.