Research On Deep Echo Cancellation Algorithm With Fused Attention Mechanism

With the progress of technology and the development of network technology,the popularity of electronic communication devices has brought a lot of convenience to people’s life and work,and the popularity of electronic communication devices has made communication between people more convenient.Communication is no longer limited to offline face-to-face communication,voice and video communication anytime and anywhere has become a trend,and real-time high-quality communication has become the pursuit of people.However,in real life,communication can be affected by the surrounding environment,equipment and network latency,and echo signals often appear in communication scenarios such as video conferencing and teleconferencing.Echo can generate interference and noise,which affects communication quality and the accuracy of speech recognition.With the popularity of deep learning in the field of machine vision,numerous scholars have proposed using deep learning to suppress and eliminate acoustic echo signals.In full-duplex communication systems,acoustic echo can degrade user experience.To address the problems of poor acoustic echo cancellation by adaptive filtering algorithms and difficulty of nonlinear acoustic echo cancellation,this paper proposes an acoustic echo cancellation algorithm based on attention mechanism,by analyzing the classical acoustic echo cancellation model based on deep learning,and improving the existing model so as to improve the model’s echo cancellation The main work and innovations of this paper are as follows:(1)A CS-BiLSTM deep acoustic echo cancellation algorithm combining attention mechanism and BiLSTM network is proposed.The BiLSTM network is constructed for the timing of speech signals,and by adding channel and spatial attention mechanisms to the BiLSTM network,the network space and channel semantic information are increased,and a new loss function is proposed by fusing root mean square error and mean absolute error to better guide the model training and improve the robustness of the model,and the ideal binary mask is used to separate and suppress the distal signal,so as to achieve the purpose of echo cancellation.The improved CS-BiLSTM network model can obtain a clear speech signal with better echo cancellation performance.Simulation results show that the proposed CS-BiLSTM algorithm significantly outperforms the reference algorithm in speech quality perception evaluation,reduces speech distortion,and achieves echo cancellation more effectively in nonlinear echo and two-way call environments,in addition,the algorithm has a simple structure and fewer model parameters.(2)An echo cancellation algorithm based on a two-branch neural network model(DBNN)is proposed,in which two branches go to learn to predict the proximal speech signal and the distal acoustic echo signal separately.The correlation between the proximal speech signal and the echo signal is obtained by introducing a feature extraction module(BSA)in the end-to-end structure to better capture the correlation features of the sequence signal.The two-dimensional residual structure is utilized as the unit component of the codec to deepen the codec hierarchy and avoid causing gradient disappearance and gradient explosion problems;and the interaction module is incorporated between two branches to improve the mutual learning modeling performance of the two-branch network for experimental comparison on different data sets.The simulation experiments show that the proposed DBNN network model achieves better results in echo cancellation with reduced distortion of the input speech signal on different data sets,and improves the intelligibility of the whole speech signal.