Audio-Based Detection of Anxiety and Depression via Vocal Biomarkers

We present a comparison of results based on the application of various model/feature combinations on the task of detecting anxiety and depression from audio signals of spontaneous speech. The adopted models comprise several different advanced deep neural networks, including CNN, LSTM, and attention networks, and are compared against traditional, shallow machine learning models. As input features we compare supra-segmental, paralinguistic feature sets against classical Mel-Frequency Cepstral Coefficients and advanced pre-trained X-vector and Wav2Vec2 features. Our models are trained based on self-assessment scores: GAD-7 for anxiety and PHQ-8 for depression. We present binary classification results for anxiety and depression separately and show that despite the noisy self-assessment labels our best model is able to achieve an unweighted average recall (UAR) of 0.60 for anxiety and 0.63 on the depression task. The result on the anxiety task almost reaches the reported self-scored GAD-7 screening reliability of 0.64. This shows that our best audio-based model can be deployed as an anxiety and depression screening tool.