Abstract: To improve the performance of speaker recognition in the condition of noise and little speech data, feature parameters were studied based on the Vector Quantization (VQ). An improved feature named WFWTC was proposed by combining extraction of Mel Frequency Cepstrum Coefficient (MFCC) with wavelet transform. After that, a new feature was established based on WFWTC and Spectral Centroid (SC). The experimental results show that the feature is feasible for speaker identification.

Key words: little speech data, speaker recognition, noisy condition, combined feature

摘要： 为了使说话人识别系统在语音较短和存在噪声的环境下也具有较高的识别率，基于矢量量化识别算法，对提取的特征参数进行研究。把小波变换与美尔频率倒谱系数(MFCC)的提取相结合，并将改进后的特征与谱质心特征进行了组合，建立了一种美尔频率小波变换系数+谱质心(MFWTC+SC)的新的组合特征参数。经实验表明，该组合特征可以有效地提高说话人识别系统的性能。

关键词: 短语音, 说话人识别, 噪声环境, 组合特征