基于Gammatone滤波器和子带能量规整的语音特征提取

为了改善传统语音特征参数在复杂环境下识别性能不足的问题,提出了一种基于Gammatone滤波器和子带能量规整的语音特征提取方法.该方法以能量规整倒谱系数(PNCC)特征算法为基础,在前端引入平滑幅度包络和归一化Gammatone滤波器组,并通过子带能量规整方法抑制真实环境的背景噪声,最后在后端进行特征弯折和信道补偿处理加以改进.实验采用高斯混合通用背景分类器模型(GMM-UBM)将该算法和其他特征参数进行对比.结果表明,在多种噪声环境中相比其他特征参数,本文方法表现出良好的抗噪能力,即使在低信噪比下仍有较好的识别效果.

Speech Feature Extraction Algorithm Based on Gammatone Filter Bank and Subband Power Normalized

In order to improve the recognition performance of the traditional speech feature parameters in complex environment,a speech feature extraction method based on Gammatone filter bank and sub-band power normalized is proposed.On the basis of the power normalized cepstral coefficients (PNCC) algorithm,this method introduces the smooth amplitude spectral envelope and the normalized Gammatone filter bank at the front end,and suppresses the background noise of the real environment by the sub-band power normalized.Feature warping and channel compensation are applied to backend processing to extract the improved feature.The GMM-UBM classifier model is used to compare the proposed algorithm with other characteristic parameters in the experiments.The results indicate that the proposed method possesses better noise robustness and preferable recognition effect than other feature parameters,under different noisy environment,even at low SNR.