Research On Speech Enhancement Method Based On Parallel Optimize Recurrent Neural Network

Speech signal is one of the important carriers for humans to transmit and express information.And the target speech will reduce the perceived quality and intelligibility after being interfered by noise,which will not only affect human subjective sense of hearing,but also reduce the accuracy of speech signal related applications.Speech enhancement,as one of the necessary technologies in the domain of speech processing,aims to reduce or remove the interference from noise,thus improving the accuracy of target speech acquisition.Traditional speech enhancement methods need to make advance assumptions about the characteristics and correlation of clean speech and interference noise,which limited the enhancement performance.Deep Neural Network solves the problem of speech enhancement by learning the mapping relation between noisy speech and clean speech,and has achieved outstanding results.The content of this thesis was based on the Recurrent Neural Network.In order to solve the problem of slow training when processing a great deal of noisy speech sequence information and further improve the enhancement performance of model,the corresponding improvement scheme is put forward.The main research work is as follows:(1)Designed and implemented a speech enhancement method based on Quasi-Recurrent Neural Network(QRNN).Long Short-Term Memory(LSTM)has limited the training speed of model because of its time dependence when dealing with a large quantity of noisy speech datasets.In order to solve the shortcomings of slow training of LSTM when processing noisy speech sequences,this paper analyzes and studies the network architecture of LSTM,improves its hidden layer unit and the input of hidden layer units,and designs a speech enhancement method based on QRNN.Compared with LSTM,the input of QRNN hidden layer is no longer independent on the output of the previous moment,and the hidden layer units is similar to the LSTM layer units,so that the model can process the noisy speech sequence in parallel while maintaining the modeling of the time change of the noisy speech sequence information.Experiments results shows that the QRNN model achieves similar enhancement performance to LSTM in speech enhancement task,and the training speed of QRNN is substantially improved compared with LSTM.(2)Designed and implemented a speech enhancement method based on QRNN incorporating Attention mechanism.Traditional neural networks cannot selectively retain useful speech sequence information,which limits the performance of speech enhancement.To further improve the performance of speech enhancement based on increasing the speed of network training,this paper combines the Attention mechanism with QRNN,designed and implemented a speech enhancement method based on QRNN incorporating Attention mechanism.This model guarantees the training speed of the network through the characteristics of QRNN parallel computing,and then assigns different weights to the sequence through the Attention mechanism to learn the target speech sequence information more selectively,in order that improving the performance of the speech enhancement model.The experimental results show that compared with QRNN network,QRNN incorporating Attention mechanism has a slightly slower training speed,but its speech enhancement performance has been significantly improved.In addition,compared with other models,it not only shortens the time of training model,but also greatly improves the performance of speech enhancement.