Speech Emotion Recognition From 3D Log-Mel Spectrograms With Deep Learning Network

• Main Authors: Hao Meng, Tianhao Yan, Fei Yuan, Hongwei Wei
• Format: Article
• Language: English
• Published: IEEE 2019-01-01
• Series: IEEE Access
• Subjects: 3-D Log-Mel; dilated CNN; residual block; center loss; BiLSTM; attention mechanism
• Online Access: https://ieeexplore.ieee.org/document/8817913/

Speech emotion recognition is a vital and challenging task that the feature extraction plays a significant role in the SER performance. With the development of deep learning, we put our eyes on the structure of end-to-end and authenticate the algorithm that is extraordinary effective. In this paper, we introduce a novel architecture ADRNN (dilated CNN with residual block and BiLSTM based on the attention mechanism) to apply for the speech emotion recognition which can take advantage of the strengths of diverse networks and overcome the shortcomings of utilizing alone, and are evaluated in the popular IEMOCAP database and Berlin EMODB corpus. Dilated CNN can assist the model to acquire more receptive fields than using the pooling layer. Then, the skip connection can keep more historic info from the shallow layer and BiLSTM layer are adopted to learn long-term dependencies from the learned local features. And we utilize the attention mechanism to enhance further extraction of speech features. Furthermore, we improve the loss function to apply softmax together with the center loss that achieves better classification performance. As emotional dialogues are transformed of the spectrograms, we pick up the values of the 3-D Log-Mel spectrums from raw signals and put them into our proposed algorithm and obtain a notable performance to get the 74.96% unweighted accuracy in the speaker-dependent and the 69.32% unweighted accuracy in the speaker-independent experiment. It is better than the 64.74% from previous state-of-the-art methods in the spontaneous emotional speech of the IEMOCAP database. In addition, we propose the networks that achieve recognition accuracies of 90.78% and 85.39% on Berlin EMODB of speaker-dependent and speaker-independent experiment respectively, which are better than the accuracy of 88.30% and 82.82% obtained by previous work. For validating the robustness and generalization, we also make an experiment for cross-corpus between above databases and get the preferable 63.84% recognition accuracy in final.