改进的DWT-MFCC特征提取算法

摘要

The wavelet transform based on DWT-MFCC is introduced into the parameter extraction of MFCC. The DWT re-places the FFT to decompose the speech signal into the wavelet coefficient with multiple sub-bands. The frequency response of the wavelet coefficient is spliced to a full spectrum directly,and obtained with the Mel filter. The improved feature extraction al-gorithm based on DWT-MFCC analyzes the wavelet decomposition process and spectrum variation of each sub-band proceeding from filtering to propose a new effective spectrum splicing method. The experimental results show that the feature extraction algo-rithm improved the recognition rate of speaker,and the cutoff characteristic of the filter and recognition rate become better with the increase of dbN length of the wavelet filter.%基于离散小波变换的美尔倒谱系数(DWT-MFCC)[1]将小波变换引入到MFCC参数的提取中,用DWT代替FFT将语音信号分解为多个频带的小波系数,并将小波系数的频率响应直接拼接为完整频谱后再通过Mel滤波器获得.改进的DWT-MFCC特征提取算法从滤波的角度分析小波分解过程及各子带频谱的变化,提出了新的有效频谱拼接方式.实验结果表明,改进的特征提取算法提高了说话人的识别率;同时,在该算法下随着小波滤波器dbN长度的增加,滤波器截止特性变好,识别率也随着增加.