| In the complex speech environment,how to build a stable and excellent speech recognition system has become an urgent problem.For this urgent need,this paper focuses on the design optimization of Chinese speech recognition system in noisy and reverberated speech environment.In this paper,spectral subtraction and ideal binary mask(IBM)algorithm are combined.To enhance the speech,first masking enhancement,then spectral de-noising.An improved IBM speech enhancement algorithm is proposed,which solves the problem of poor speech processing effect of classical speech enhancement algorithm under strong noise interference.The experimental results show that the improved IBM algorithm has better speech processing effect than the existing five classical speech enhancement algorithms when the noise intensity coefficient is greater than 0.008 and reverberation is included.But when the noise intensity coefficient is less than 0.008 and reverberation is included in the speech,the improved IBM algorithm is not as good as the original IBM algorithm.Based on this,an adaptive speech enhancement algorithm based on noise spectrum estimation is proposed.The idea of the algorithm is to estimate the noise spectrum and judge the noise intensity coefficient before speech enhancement.If the noise intensity coefficient is less than 0.008,the IBM algorithm is selected for processing.On the contrary,the improved IBM algorithm is used for speech enhancement.The experimental results show that the proposed algorithm has better performance than other algorithms in the whole noise and reverberation condition.In Chapter 4,an adaptive Chinese speech recognition system based on noise spectrum estimation is built.A speech enhancement module is added to the front-end of the existing speech recognition system,in which the convolutional neural network and connectionist temporal classification technology are used to build the acoustic model and the maximum entropy markov algorithm is used to build the language model,and the adaptive IBM speech enhancement processing module based on noise spectrum estimation is embedded to build the speech recognition system.The working principle of the system is that in speech recognition,the noise spectrum is estimated and the noise intensity is judged first,then the speech enhancement algorithm is selected adaptively to enhance the speech,and finally the speech recognition is carried out.The experimental results show that the speech recognition system built in this paper has higher speech recognition accuracy for noisy and reverberated test set speech.At the same time,experiments show that the error rate of Chinese character recognition is much higher than that of Pinyin recognition.Based on this,in the fifth chapter,a neural network language model based on self-attention mechanism is built by combining attention mechanism with deep learning technology.The experimental results show that compared with the maximum entropy markov language model based on probability spectrum and the neural network language model built by CBHG module composed of one-dimensional convolution group,high-speed network and two-way gated recurrent unit,the language model built in this paper has higher accuracy of pinyin to Chinese characters.At the same time,the language model is embedded into the speech recognition system in Chapter 4,which also achieves higher speech recognition accuracy. |
