一种基于噪声对消与倒谱均值相减的鲁棒语音识别方法

A robust speech recognition method by combining noise cancelling and cepstral mean subtraction

提出一种基于语音增强算法的噪声鲁棒语音识别方法.在语音识别预处理阶段,通过噪声对消语音增强法来抑制噪声提高信噪比.然后对增强语音提取Mel频段倒谱特征参数,并在倒谱域应用倒谱均值相减处理来补偿增强语音中的失真成分和剩余噪声.实验结果表明,在低信噪比(-12～0 dB)条件下,该方法对于数字语音识别具有较好的识别率,其性能明显优于基本的Mel频段倒谱参数识别器、传统的谱减法和噪声对消语音增强法.

A noise resistant speech recognition method based on a speech enhancement algorithm was implemented. First, it obtains the denoised speech, with significant SNR （signal-to-noise ratio） improvement, by applying adaptive noise cancelling （ANC） to the pre-treatment stage of speech recognition. Then Mel-frequency cepstral coefficients（MFCC） are computed from the enhanced speech. Then cepstral mean subtraction （CMS） is used to compensate for components of distortion and the residual noise of the enhanced speech in the cepstral domain. When speech samples have a low SNR, ranging from 0 to 12 dB, experimental results indicate that the proposed method performs better than a standard MFCC recognizer, conventional spectral subtraction （SS） and the ANC speech enhancement for digital speech recognition.