Research on Real-Time Speech Enhancement Models Based on Feature Compression
AA novel real-time speech enhancement algorithm based on feature compression is proposed,which combines signal processing methods and neural network methods. On the one hand,it reduces the complexity of the model,and on the other hand,it effectively suppresses noise. The algorithm uses traditional speech signal processing methods to process and compress features,in order to extract more implicit features from speech. Then,a neural network is used to learn the extracted features,and finally,the corresponding pure speech signal prediction is obtained. The proposed algorithm can achieve real-time processing (one frame in,one frame out),and the objective evaluation indicators and mean opinion score of speech quality are close to or even exceed the baseline algorithm when the complexity of the baseline model is lower,demonstrating good noise reduction performance. In terms of model parameter scale,the model parameters of the proposed algorithm are much smaller than those of the baseline algorithm.
noise suppressionneural networkssignal processingreal time processingfeature compression
万方数据
维普
