Parametric Resynthesis Speech Enhancement Algorithm Based On Feature Learning

The speech enhancement system should completely eliminate the noise for the speech polluted by noise without reducing the speech quality and speech intelligibility under ideal circumstances.However,the current speech enhancement system will lose part of the speech information parameters while removing the noise.The parametric resynthesis speech enhancement algorithm integrates the traditional speech enhancement algorithm and TTS(Text to Speech,TTS),and experiments prove the superiority of the combined system in enhancing performance,it reappears all parts of the Speech signal,even the Speech information hidden in the noise,to obtain a high enhancement performance,but there are some shortcomings in the system so that the enhancement performance of the system is affected.In this paper,the problems in traditional parameter resynthesis are improved,and the specific research work is as follows:(1)Because the parametric resynthesis speech enhancement system uses single acoustic features and non-neural network vocoder,the enhanced performance of Parametric Resynthesis speech enhancement system will be affected.In order to solve the above problems,paper proposes a parametric resynthesis speech enhancement algorithm based on multi-feature fusion,first of all,a variety of acoustic features are integrated through the attention mechanism,then uses comprehensive acoustic features of noisy speech instead of the single acoustic features to complete prediction,as a result,the system can retain more information about clean speech;On this basis,the neural network vocoder Wave Net vocoder is used to synthesize high-quality clean speech.Finally,the experimental comparison verifies that the proposed algorithm can effectively improve the speech quality.(2)In the parameter resynthesis speech enhancement system,the acoustic features of noisy speech are extracted manually by signal processing method,and the acoustic features of clean speech is predicted through Deep Neural Networks DNN.The above behaviors will result in the loss of the spatial structure and temporal information of the original noisy speech,so,the enhanced performance of system will be affected.Paper proposes a parameter resynthesis speech enhancement algorithm based on convolutional gated cyclic network to improve the performance of the system,model consists of two layers,the first layer extracts speech features by the convolution neural network with attention mechanism to retain spatial structure information of input speech,and the second layer learns the temporal information of speech and predicts the acoustic features of clean speech by Inner Attentive Bidirectional GRU network-bi GIAG.Then,clean speech will be synthesized by vocoder.Finally,experiments prove that the proposed method can effectively improve the enhancement performance of the original system.