基于atan-LMS与PLPC相结合的稳健语音识别

Noise Robust Speech Recognition Based on atan-LMS and Perceptual Linear Prediction Coefficients

将一种改进的变步长算法(atan-LMS)应用于自适应滤波器,并将之与感知线性预测倒谱技术(PLPC)相结合,简记为atan-LMS+PLPC。该方法通过优化步长和误差之间的非线性关系,使降噪效果更加显著,然后对增强语音提取感知线性预测倒谱系数(PLPC)作为特征参数。试验结果表明,基于atan-LMS+PLPC的系统其识别性能要比基于Sigmind函数(SVSLMS+PLPC)优越,并且信噪比越低识别率提高幅度越大。

A method named atan-LMS＋PLPC by combining adaptive filtering based a new variable step size LMS and Perceptual Linear Prediction Coefficients （PLPC） is proposed. It can get better result by optimizing the relation between step size and error. And then, PLPC is computed. According to the experimental result, ant-LMS performs better for speech recognition with Sigmind function （SVSLMS） and PLPC. Moreover, recognition rate increases faster with the signal-to-noise decreasing.