基于FMFCC和HMM的说话人识别

Speaker Recognition Based on FMFCC and HMM

美尔频率倒谱系数(MFCC)是说话人识别中常用的特征参数,而语音信号是非平稳信号,MFCC并不能很好的反映语音的时频特性。针对这一缺陷,为了提高说话人的识别率,结合新的时频分析工具分数傅立叶变换(FRFT)。将MFCC推广到分数形式,得到分数美尔频率倒谱系数(FMFCC),用以表征语音信号的特征;并利用可分性测度验证了特征参数的有效性;通过建立20个不同说话人的FMFCC特征库,采用隐马尔可夫模型(HMM)对说话人进行仿真识别。仿真结果表明,在合适的变换阶次下,说话人的平均识别率可达93%以上。

Mel frequency cepstral coefficient （MFCC） is a frequently - used characteristic in speaker recogniton. In evidence, speech are non - stationary signals, the time - frequency characteristic of speech is not clearly expressed through MFCC. Thus, in the calculation of MFCC parameter, fractional Fourier transform （FRFT） is adopted to replace discrete Fourier transform. Then fractional Mel frequency cepstral coefficient （FMFCC） is acquired, and the effectivity of the parameter is verified. Finally, the Hidden Markov Model （HMM） of 20 different speakers is estab- lished, and speaker identification is performed. The simulation shows that in different transform orders, the average of right speaker recognition rate is up to 93%.