基于改进的隐马尔可夫和神经网络混合模型的语音识别

Hybrid speech recognition based on improved hidden markov model and neural network

研究了一种非齐次隐马尔可夫模型(Inhomogeneous Hidden Markov Model),然后将自组织特征映射神经网络与这种非齐次隐马尔可夫模型相结合,训练出抗噪声的HMM模型,并应用该混合模型进行语音识别。实验结果表明,该模型适合于对噪声背景下的语音进行识别。该模型具有更好的抗噪鲁棒性,在信噪比较低的情况下(5dB-10dB),识别率可以提高5％左右。

The inhomogeneous-HMM is studied, and the Self-Organizing Feature Mapping neural network-SOFMNN and an improved inhomogeneous-HMM are combined to train the antinoise HMM. The model trained by this method is used in speech recognition experiments. Experimental results show this model has better noise robustness. In the condition of low SNR （5dB-10dB）, the correct recognition rate increased about 5%.