Study On Far-Field Real-Time Acoustic Echo Cancellation Algorithm

In recent years,wireless voice communication technology has made remarkable progress and various voice interaction devices and softwares have also emerged.The quality of the voice during the call will directly affect the user’s listening experience.Acoustic echo is the major factor of interference with speech quality.In a call scenario with a high signal-to-noise ratio,the traditional acoustic echo cancellation algorithms can greatly improve the voice quality.However,when people are far away from the devices or there are various noises in the sound field environment,the performance of the acoustic echo cancellation algorithms are confronted with huge challenges.In order to maintain speech communication qualities in these complex scenarios,this thesis proposes a real-time subband acoustic echo cancellation algorithm combining dual-filter structure and fast Kalman algorithm.And a post-filtering algorithm based on neural network is improved to suppress residual echo and noise.This thesis firstly studies the acoustic echo cancellation algorithm with low computational complexity and robust double-talk performance.In this thesis,the sub-band processing method is used to realize the fast Kalman filter,which can deal with the problem that the filter order required in the far-field scene is too long which causes the algorithm to be too computationally intensive to satisfy real-time requirements.Compared with the most commonly used normalized least mean square algorithm in acoustic echo cancellation,the fast Kalman algorithm has slightly higher computational complexity but better convergence performance.In order to ensure the echo cancellation performance during double-talk,a dual-filter structure is adopted in this thesis.The echo path is estimated in real-time by a continuously adaptively updated background filter.It is controlled by the transfer logics to transfer the tap coefficients to the non-autonomously updated foreground filter to obtain the final output result in a high-interference environment.The key to the performance of the dual-filter is the design of the transfer logics.This thesis proposes a set of transfer logics that are more robust than traditional methods,controlling the bidirectional transfer of tap coefficients for both the background and foreground filters.Experiments show that the algorithm proposed in this thesis is superior to the comparison algorithms in terms of echo suppression capability,convergence speed and double-talk mode processing,and the new transition logics improves the robustness of the algorithm.When the caller is in a noisy speech communication scenario,ambient noise will affect the call quality.In addition,speech that has undergone linear echo cancellation often has a small amount of residual echo,which needs to be further processed by a post-filtering module.This thesis improves a real-time post-filtering algorithm based on recurrent convolutional network.Experimental results show that the improved post-filtering algorithm can better suppress residual echo and noise,and the improvement of subjective and objective speech quality indicators is greater than that of traditional methods.