基于Teager能量算子(TEO)非线性特征的语音情绪识别

目的 探索识别汉语语音情绪的有效识别特征.方法 采用基于Teager能量算子（TEO）的非线性特征,通过马尔可夫模型法（HMM）,从汉语语音中识别平静和生气、欢快、悲伤4种情绪.结果 文本有关时,5个非线性特征：基于频域TEO的Mel倒谱系数（nonlinear frequency domain Mel,NFD_Mel）、基于幅频特性的Mel倒谱参数（amplitude and frequency property Mel,AF_Mel）、基于微分幅频特性的Mel倒谱参数（amplitude and frequency property Mel of differential,DAF_Mel）、基于幅度调制的子带倒谱参数（AM-based SBCC,AM_SBCC）及基于幅频调制的子带倒谱参数（AMFM-based SBCC,AMFM_SBCC）的情绪识别性能全部高于Mel频率倒谱参数（Mel-scaled cepstrum coefficients,MFCC）.文本无关时,NFD_Mel、AF_Mel、DAF_Mel的识别率高于MFCC,AM_SBCC、AMFM_SBCC的情绪识别率低于MFCC.结论 结合非线性TEO的识别特征NFD_Mel、AF_Mel、DAF_Mel可有效提高情绪识别性能.

Emotion Recognition of Mandarin Speech Using Nonlinear Features Based on Teager Energy Operator(TEO)

Objective To study effective features of mandarin speech which represent different emotions in speech. Method Using nonlinear features based on Teager energy operator (TEO) and adopting hidden markov model(HMM) recognizing method , neutral state and 3 emotional status( happiness, anger and sadness) were recognized from the mandarin speech database. Result In comparison with Mel-scaled cepstrum coefficients(MFCC), while text-dependent, improvements of recognition capacity were obtained when using all 5 nonlinear features: nonlinear frequency domain Mel(NFD-Mel),amplitude and frequency property Mel(AF-Mel),amplitude and frequency property Mel of differential(DAF-Mel), AM-based SBCC(AM-SBCC) and AMFM-based SBCC(AMFM-SBCC). While text-independent, the performance of emotion recognition was improved by using NFD-Mel , AF-Mel and DAF-Mel , but deteriorated by using AM-SBCC and AMFM-SBCC. Conclusion The results of recognition demonstrated that the nonlinear features based on TEO, when using NFD-Mel, AF-Mel and DAF-Mel, can increase emotion recognition capacity.