关键词: 语音情感识别, 多层次通道融合, wav2vec 2.0, 卷积神经网络（CNN）

Abstract: Speech emotion recognition is key to the emotional cognitive ability of machine and crucial for improving human-machine interaction quality. However, most of the existing studies focus on the analysis of shallow features, ignoring the advantages of multi-feature fusion. Additionally, the limited size of data samples affects models?? generalization capability, thereby resulting in suboptimal accuracy in speech emotion recognition. In order to further improve the accuracy of speech emotion recognition, a method based on data augmentation and multi-level channel fusion is proposed. Firstly, the original speech is enhanced using Gaussian white noise, pitch shift, and mixed processing, to improve the model??s robustness. Secondly, based on the wav2vec 2.0 model and CNN model, a multi-level parallel channel network is proposed. The first channel uses the wav2vec 2.0 model as the backbone network to learn deep representations of speech data, followed by two layers of convolutional CNN model for computation. The second channel extracts shallow emotion features of speech as input and uses a five-layer convolutional CNN model to learn the shallow representations, enabling more comprehensive analysis of both deep and shallow representations in speech data. Finally, the representations from both channels are fused, forming a multi-level speech emotion feature system that combines deep and shallow features. The proposed model is tested on the RAVDESS and CASIA datasets, achieving accuracy of 94.38% and 98.75%, respectively, validating the effectiveness of the proposed approach.

Key words: speech emotion recognition, multi-level channel fusion, wav2vec 2.0, convolutional neural network (CNN)