汉语数码语音识别中一种新的抗噪声特征参数

New Anti-noise Feature for Mandarin Digital Speech Recognition

为了提高中小词汇量语音识别系统在噪声环境下的识别性能,以10个汉语数码语音为对象,利用汉语数码语音信号区别于噪声信号的准周期特性,提出了一种汉语数码语音频谱包络峰值特性的提取方法,首先用基频对语音频谱采样得到由谐波值构成的包络以提高信噪比,然后再对所得包络进行峰值提取得到汉语数码语音的峰值特征。实验结果表明,在信噪比大于5dB时,用该方法得到的峰值特征具有一定的抗噪性。

A new peak feature selection method for Mandarin digital speech recognition is introduced.This method isbased on the basis of different period features between Mandarin digital speech and noise and divided into two steps.The first step is to sample the frequency spectrum of speech by fundamental frequency to get its envelope andincrease its SNR. The second step is to select the peak feature of the envelope. The experimental results indicate thatthe peak feature of the envelope has an anti-noise character when SNR > 5dB.