关键词: Mel尺度, Hilbert边际能量谱, 边际能量谱特征, 情感识别

Abstract: Emotional feature extraction plays an important role in speech emotion recognition. Due to the limitations of traditional signal processing methods, traditional phonetic features, especially in terms of frequency domain features, are unable to reflect precisely phonetic emotional characteristic, which leads to a low emotion recognition rate. This paper proposes a new method. Firstly, Hilbert-Huang Transform（HHT） is used in order to process speech signal, thus to obtain Hilbert marginal energy spectrum. Then, a comparison and relative analysis based on Mel-scale is carried out, afterwards a new array of emotional features are obtained, which consists of Mel-Frequency Marginal Energy Coefficient（MFEC）, Mel-frequency Sub-band Spectral Centroid（MSSC）and Mel-frequency Sub-band Spectral Flatness（MSSF）. Finally, the five kinds of speech emotion namely sadness, happiness, boredom, anger and neutral are recognized by using the Support Vector Machine（SVM）. The experimental results show that the new emotional features extracted by this method have better recognition performance.

Key words: Mel-scale, Hilbert marginal energy spectrum, marginal energy spectrum feature, emotion recognition