中文语音情感常用特征识别性能分析

The performance analysis of feature in Chinese Speech Emotion Recognition

语音情感识别对于实现人机交互具有重要的应用价值。语音情感识别中,情感特征的选取与组合对于情感识别的准确率影响巨大。已有研究中,情感特征对识别率的贡献停留在定性分析中,未有定量的描述,不利于情感识别中特征的选择。本文针对中文语音情感识别中的常用特征进行定量分析,通过不同的情感特征进行组合,采用支持向量机进行分类,得到各情感特征对识别的贡献率。实验结果表明,单个特征中,梅尔倒谱系数贡献率最高,达到了78%;特征组合中,特征越多对识别率贡献越大。

The Speech Emotion Recognition is a vital technique for human-computer interaction. The selection and fusion of emotion feature is a key impact factor to accuracy of recognition. In existing study, the contribution of emotion feature to recognition rate is still stay on perception analysis, the quantitative analysis is rarely reported. However, the perception analysis is unhelpful for the feature selection in applications. This paper proposes a quantitative analysis method for helping select the emotion feature on Chinese Speech Emotion Recognition. In the proposed method, the research obtains the contribution of features for emotion recognition by different feature combination on Support Vector Machine classifier. The experiments show that Mel cepstrum coefficients is highest and reaches to 78% when Using single feature for recognition, and more feature and more high contribution when using feature combination.