This study explores gender and race bias in machine learning models that use speech to detect mental health outcomes, namely anxiety and depression. The authors examined where bias can come from, both in the acoustic features extracted from speech and in the labels used to train the models, as well as in the models’ decisions.
They also tested two ways of reducing bias: using only the speech features that carry the least demographic information, and transforming the data to remove traces of demographic information while keeping what matters for detecting the mental health condition. Performance varied across demographic groups, and the results of these approaches were mixed, highlighting the need for careful design so that speech-based tools work well for everyone.
Learn more about this study here: https://doi.org/10.3389/fdgth.2024.1351637
Reference
Yang M, El-Attar AA and Chaspari T (2024) Deconstructing demographic bias in speech-based machine learning models for digital health. Front. Digit. Health 6: 1351637.