Improvement on Speech Depression Recognition Based on Deep Networks

摘要

To reduce the burden of clinicians diagnosing a large number of depressive symptoms, the field of artificial intelligence researchers are increasingly interested in designing automatic recognition systems for depression. Depressed patient have different speech signal from normal people. Here, we present a deep model, Depression AudioNet, which encodes depression-related features in the vocal tract and provides a more comprehensive audio representation. Firstly, the Mel-frequency cepstral coefficients (MFCCs) were extracted from raw audio data. Secondly, the robust emotions features were acquired by Multiscale Audio Delta Normalization (MADN), which is a data processing algorithm we proposed. Finally, the MFCCs and the emotions features of two adjacent segments of local audio were fed into the Depression AudioNet in turn to train the network. This method solves the problem of less training data and low precision by increasing the length information of the sample without reducing the number of samples. Experiments are conducted on AVEC2014 dataset, and the results shows that the proposed method is more effective and accurate than the existing speech depression recognition algorithms.