一种基于HMM/ANN的混合语音识别系统的设计

A Hybrid Speech Recognition System Based on HMM/ANN

文中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,训练出抗噪声的HMM模型。试验表明,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(3dB—15dB),识别率比传统CDHMM模型有明显的提高。

This paper proposed a hybrid model method combining Self-Organizing Feature Mapping neural network with Hidden Markov Model to train noise adapting HMM. The model trained by this method is conformable to recognize the speech in noisy environment. Compared with the traditional CDHMM and the CDHMM trained by additive noise into speech, this model have better noisy robustness. In the condition of SNR is low(2dB-12dB), the correct recognition rate increased distinctly.