计算机工程与应用 ›› 2014, Vol. 50 ›› Issue (3): 198-202.

• 信号处理 • 上一篇    下一篇

基于HHT倒谱系数的说话人识别算法

杜晓青,于凤芹   

  1. 江南大学 物联网工程学院,江苏 无锡 214122
  • 出版日期:2014-02-01 发布日期:2014-01-26

Speaker recognition algorithm based on HHT cepstrum coefficient

DU Xiaoqing, YU Fengqin   

  1. School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2014-02-01 Published:2014-01-26

摘要: 针对LPCC只反应语音静态特征且不能突出其低频局部特征问题,提出一种以HHT倒谱系数为特征的说话人识别算法,HHT的经验模态分解使语音的低频局部特征得到更好的描述,Hilbert变换能够刻画语音动态特性,改进了LPCC的不足。用经验模态分解将语音分解为一系列固有模态函数分量并做Hilbert变换求得Hilbert边际谱,计算总边际谱的对数功率谱并做DCT得13维倒谱系数,将此特征送入高斯混合模型进行说话人识别。仿真实验结果表明,基于HHT倒谱系数的说话人识别算法,相较LPCC识别率提高了12.59%,但特征提取时间增加了19.27 s。

关键词: 说话人识别, 希尔伯特黄变换(HHT), 倒谱系数

Abstract: According to the problem that LPCC only reacts speech signal static characteristics and can not describe the low frequency local characteristics of speech signal well, a new speaker recognition algorithm based on HHT cepstrum coefficient is proposed. The low frequency local characteristics of the signal can be described better by the empirical mode decomposition of HHT. The dynamic characteristics are reacted by the Hilbert transform, improving the LPCC deficiencies. Speech signal is decomposed into intrinsic mode components using empirical mode decomposition. Hilbert transform is done for each component to get the Hilbert marginal spectrum. The logarithmic power spectrum of total marginal spectrum is calculated and then done the DCT to get 13-dimensional cepstrum coefficient. The feature is sent into the gaussian mixture model to do speaker recognition. Simulation results demonstrate that compared to the LPCC, the HHT cepstrum coefficient gets a higher recognition rate. Recognition rate is increased by 12.59%, but feature extraction time is increased by 19.27 s.

Key words: speaker recognition, Hilbert-Huang Transform(HHT), cepstrum coefficient