渐进学习语音增强方法在语音识别中的应用

Application of Progressive Learning Speech Enhancement Approach to Speech Recognition

在语音识别实际应用中,带噪语音信噪比的复杂性会造成识别难度增大,导致语音识别系统性能下降.本文将渐进学习语音增强方法应用于语音识别,以取代传统语音识别中使用的基于深层神经网络的语音增强方法.本文使用渐进学习语音增强方法在识别模型前端进行降噪预处理,然后再作识别,以更好地提升语音信噪比,进而提高系统性能.首先使用渐进学习方法训练一个深层神经网络.然后,将语音经过这个渐进学习深层神经网络作增强.最后,将渐进学习深层神经网络增强后的语音经过语音识别模型作识别.通过实验验证,本文使用的渐进学习语音增强及识别方法,相对于传统语音增强及识别方法,在识别准确率上有10.28%的相对提升.

In the practical applications of speech recognition, the SNR （ Signal to Noise Ratio ） complexity of noisy speech will resultin the reduction of system performance. In this paper ,we used the PL （ Progressive Learning ） speech enhancement method in the ap-plication of speech recognition,to replace the traditional speech enhancement method based on DNN（ Deep Neural Network}. In orderto improve the SNR of speech and the system performance,the speech enhancement PL-DNN is used for noise reduction,in the front-end of speech recognition model. Firstly, the progressive learning method is used to train a PL-DNN. Then the speech is enhanced withthis PL-DNN. Then the enhanced speech is used for recognition. From experimental results, the PL-DNN method used in this paper canoutperform the traditional enhancement and recognition based on DNN, with a recognition accuracy relative improvement of 10.28 %.