摘要
提出了一种语音识别线性预测分析方法 :基于谱自相关和频率抽样获得谱包 ,即由归一化频率估计谱包 ,此谱包规定在 Mel频率级 ;再由语音信号谱包估计抽样自相关 ,用 IDFT提取抽样自相关估计。从抽样自相关的结果 ,最终获得谱包倒谱系数。HMM识别试验显示 :谱包倒谱系数与其他算法相比较 ,在低信噪比时 ,识别率可提高 1 0 %以上 ,识别性能明显提高 ,在噪声环境下也能达到好的识别效果。
A linear predictive analysis method of speech recognition for estimating sample autocorrelation from the speech signal spectral envelope is proposed based on spectral autocorrelation. To obtain spectral envelope from estimating frequency samples a frequency normalization can be applied to the estimated spectral envelope. The spectral envelope is the mel frequency scale and IDFT is used to extract the estimate of sample autocorrelations. The cepstral coefficients are obtained from sampling autocorrelation results. HMM experiments show that cepstral coefficients improve the performances of the recognizer at low R SN . The recogniton rate is improved more than 10% and it works well in noise environments.
出处
《数据采集与处理》
CSCD
2004年第4期421-424,共4页
Journal of Data Acquisition and Processing
基金
河南省自然科学基金 (0 41 1 0 1 0 1 0 0 )资助项目。
