基于连续HMM的孤立语音鲁棒性识别方法

A Robust Recognition Method of Isolated Words Based On Continuous HMM

对于基于连续隐马尔可夫模型（CHMM）的语音识别系统，为了提高系统在环境噪声下的鲁棒性，本文提出了一种能有效抑制加性平稳噪声和通道卷积噪声的相对自相关序列的Mel倒谱参数（RAS_MFCC＋△RAS_NFCC），进行特征参数级的去噪，明显地改善了系统的噪声鲁棒性。为了进一步提高系统在低信噪比语音时的识别性能，我们采用了CHMM的混合语青训练法，获得了对各种信噪比语音都具有很强适应性的CHMM参数。实验证明。

In order to improve the lloise robustness of speech recognition system based on continuous Hidden MarkovModels(CHMM), Mel-frequency cepstrum coefficient of relative auto-corre1ation sequence (RAS_MFCC+ △ RAS_MFCC) isproposed to suppress additive stationary and convolution noise in the feature space. This method can be used to enhance thesystem robustness effectively as to improve the recognition rate in the low signal-to-noise ratio environment. We also adopt amethod called hybrid speech training of CHMM where the trained CHMM parameters have better adaptive ability to theenvironment with various signal-to-noise ratios. Experiments show that this method combining with noise suppression infeature space and compensation at system mode1 level can greatly improve system'S recognition performance in noisecontaminated environments.