基于多分形谱及特征优选的说话人识别系统

Speaker Recognition System Based on Multifractal Spectrum Feature and Characters Selection Policy

摘要: 语音是复杂的非线性信号，这使得基于线性理论的传统说话人识别系统性能难以进一步提高。结合语音特点，基于小波极大模方法(Wavelet Transform Modulus-Maxima Method, WTMM)，提出一种语音多分形谱特征(Multifractal Spectrum Feature, MSF)提取方法，并将语音多分形谱特征与传统特征结合用于说话人识别，实验表明，在短语音说话人识别中，6维MSF与LPC结合，误识率相比单独使用LPC降低了6.4个百分点；而MSF与MFCC、LPC组合，误识率降至1.2%左右。采用贪婪策略对说话人识别的特征进行优选，从101维特征中优选出13维特征用于识别，实验结果表明优选后的特征参数能有效降低系统误识率，提高识别速度，误识率最低降至1.6%，识别时间减少约86%。

Abstract: Speech is one kind of complicated non-linear signal, so traditional speech or speaker recognition system based on the linear theory is difficult to be further improved. In this paper, a new method based on the WTMM (wavelet transform modulus-maxima method) is proposed, which can facilitate the extraction of speech signals in the multifractal spectrum feature (MSF). The multifractal spectrum feature combined with the traditional linear features can obviously enhance performance of speaker recognition system. Experiment results show that 6-dimensional MSF combined with 13-dimensional MFCC and 16-dimensional LPC make error rate decrease to 1.2% in short speech speaker recognition. Then greedy algorithm is used to select 13 dimensional features from 101-dimensional features set. The experiment results show that the optimal feature selective method can eliminate disturbance of other redundant features, and obviously reduce the error rate, and improve the computational speed. The error rate decreases to 1.6%, and computation time decreases about 86%.