Convolutional Neural Network with Spectrogram and Perceptual Features for Speech Emotion Recognition

Convolutional neural network (CNN) has demonstrated a great power at mining deep information from spectrogram for speech emotion recognition. However, perceptual features such as low-level descriptors (LLDs) and their statistical values were not utilized sufficiently in CNN-based emotion recognition. To solve this problem, we propose novel features to combine spectrogram and perceptual features in different levels. Firstly, frame-level LLDs are arranged as time-sequence LLDs. Then, spectrogram and time-sequence LLDs are fused as compositional spectrographic features (CSF). To fully utilize perceptual features and global information, statistical values of LLDs are added in CSF to generate rich-compositional spectrographic features (RSF). Finally, the proposed features are individually fed to CNN to extract deep features for emotion recognition. Bi-directional long short-term memory was employed to identify emotions and the experiments were conducted on EmoDB. Compared with spectrogram, CSF and RSF improve the unweighted accuracy by a relative error reduction of 32.04% and 36.91%, respectively.