結合非線性動態特徵之語音情緒辨識

Speech Emotion Recognition via Nonlinear Dynamical Features

本研究採用機器學習法對語音情緒辨識進行探討。除一般常被採用之語音特徵，如音高、共振峰、能量以及梅爾倒頻譜係數之外，研究中加入了夏農熵和曲率指標（curvature index）兩項非線性特徵，再利用費雪鑑別比與基因演算法搭配的方式進行特徵挑選。最後使用支持向量機分類器，對柏林語音情緒資料庫進行情緒分類分析。在加入非線性特徵後，男性及女性之情緒辨識率分別為88.89%及86.21%。

This study is focus on speech emotion recognition through machine learning method. We add two nonlinear dynamical features: Shannon entropy and curvature index, of each frame other than the traditional features such as pitch, formant, energy, MFCCs. After feature extraction, Fisher discriminant ratio and Genetic algorithm were applied in order to reduce the number of features. We use SVM classifier and cross validation method to discriminate seven emotions in Berlin emotion database. The analyzed results after adding of the nonlinear features show that the emotion recognition rates were 88.89% and 86.21% for male and female, respectively.