Single Channel Speech Enhancement Algorithm Based On Attention Mechanism DNNs

A single channel speech enhancement method based on attention mechanism and multiple deep neural networks (DNNs) is proposed to address the problems of neglecting the correlation between features and inaccurate noise suppression in speech enhancement. The attention mechanism model consists of a masking estimation network and an auxiliary training network, fully considering the importance of speaker adaptive speech representation. Firstly, train an auxiliary neural network (ANN) that is independent of the speaker model; Then, the speech sequence summarization method is adopted and channel attention is introduced to extract speaker adaptive information based on ANN. This information is used as an embedding vector input to the hidden layer of the masking estimation network, and the activation output of this layer is scaled with bias; Finally, output masking is used to achieve speech enhancement. Experiments were performed on the TIMIT and NOISEX-92 datasets, and the results showed that the model performed better when introducing speaker adaptive information under low signal-to-noise ratio conditions. Compared with single masking training networks and other baseline methods, the proposed algorithm model showed significant improvements in PESQ and STOI indicators.