基于CDHMM/SOFMNN噪声背景下的语音识别方法

针对噪声背景下传统语音识别系统识别率较低的问题,提出了一种将自组织特征映射神经网络(SOFMNN)与隐马尔可夫模型(HMM)相结合的方法,训练出适应噪声的混合模型.该模型适合于对噪声背景下的语音进行识别.同传统的CDHMM模型以及直接在语音中加入加性噪声训练出的CDHMM模型相比,该模型具有更好的抗噪鲁棒性,在信噪比较低的情况下(2~12 dB),识别率比传统CDHMM模型有明显提高.

Speech recognition system based on CDHMM/SOFMNN in noisy environment

Aiming at the question of the low recognition rate of the traditional speech recognition system, this paper proposes a hybrid model method combining self-organizing feature mapping neural network with hidden Markov model to train a noise adapting model. The model trained by this method is suitable to recognize the speech in noisy environment. Compared with the traditional CDHMM and the CDHMM trained by additing noise into speech, this model has better anti-noise robustness. Under the condition of low SNR (2~12dB),the correct recognition rate increases significantly.