关键词: 语音情感识别, 深度学习, 多特征融合, 数据增强, 长短时记忆神经网络, 简化加性注意力机制

Abstract: Aiming at the problems of incomplete representation of speech information by single speech emotion features and low utilization rate of speech features by the model, a multi-feature fusion speech emotion recognition method based on SAA-CNN-BiLSTM network was proposed. In this method, noise, volume and sound speed boosters were introduced to enhance the data, enabling the model to learn diverse data features. Multiple features including fundamental frequency, time domain and frequency domain were fused to comprehensively express emotional information from different perspectives. Based on BiLSTM network, CNN was introduced to capture the spatial correlation of the input data and extract more representative features. A simplified additive attention mechanism was constructed, and the explicit query keys and query vectors were simplified, so that the calculation of attention weights does not depend on specific query information. Features of different dimensions can be correlated and influenced based on the attention weights, and the information between features can be interacted and fused, thus improving the effective utilization rate of features. The experimental results show that this method achieves accuracies of 87.02%, 82.59%, and 73.13% on the EMO-DB, CASIA, and SAVEE datasets respectively. Compared with the baseline methods such as IncConv, NHPC-BiLSTM, and DCRNN, it improves the performance by 0.52-9.80 percentage points, 2.92-23.09 percentage points, and 3.13-16.63 percentage points respectively.

Key words: speech emotion recognition, deep learning, multi-feature fusion, data augmentation, long short-term memory neural network, simplified additive attention mechanism