Research On Acoustic Echo Cancellation Method Based On Deep Neural Network

In everyday communication,due to network transmission,the surrounding acoustic environment,voice playback and acquisition equipment,etc.,there is usually an echo in the caller’s voice,and the high latency of acoustic echo reduces the recognition of useful information,causing serious interference to people’s sense of experience during the call.Acoustic echo cancellation technology is designed to eliminate acoustic echo from communication devices and applications in order to improve the voice quality of people during calls and to provide a good user experience.Traditional acoustic echo cancellation algorithms based on adaptive filters are usually suitable for eliminating echoes in the single-speaker state.With the rapid development of deep learning in recent years,acoustic echo cancellation techniques based on deep neural networks have been well developed,but the model structure is developing in a deep and complex direction,and the processed audio has a small amount of nonlinear acoustic echo residuals.Therefore,this paper aims to improve the acoustic echo cancellation performance while reducing the number of neural network parameters and eliminating the residual nonlinear acoustic echoes.(1)This paper proposes an acoustic echo cancellation method based on a residualbi-directional long short-term memory network.The residual structure is used to learn different levels of acoustic echo abstraction and deep temporal feature information,and the bidirectional long and short term memory network is used to improve the model’s ability to learn the contextual relationship of the speech signal for the temporal nature of the speech signal.And by replacing the conventional convolution with a depthwise separable convolution,the number of parameters is reduced to a greater extent,achieving better results than traditional echo cancellation methods.(2)In this paper,an acoustic echo cancellation method combining the optimized training objective and neural network is proposed.In order to pay attention to the influence of the characteristics of the amplitude spectral similarity of speech signals on the performance of echo cancellation,by making full use of the amplitude spectral similarity between the proximal microphone signal,the proximal speech signal and the acoustic echo,the mutual relationship number is introduced and the accurate binary mask is constructed as the training target of the model,and the proposed accurate binary mask guides the model training to achieve better echo cancellation results than other masking methods.In this paper,we propose an acoustic echo cancellation method combining Kalman filter and neural network.To address the problem that residual echoes still exist in the audio processed by the echo cancellation algorithm,the Kalman filter is used to process some of the acoustic echoes and noise,thus reducing the training pressure on the neural network,and then the neural network is used to eliminate the residual echoes,combining the advantages of both and achieving a good evaluation score.Since the linear adaptive filter is slow in processing audio,a combined time-frequency domain network is proposed for acoustic echo cancellation,where the neural network in the frequency domain can learn more accurate spectral information of the speech signal and the neural network in the time domain performs end-to-end processing to learn more accurate time-domain waveforms of the speech signal to further eliminate residual acoustic echoes.The combined time-frequency domain approach effectively improves the ability of the model to eliminate acoustic echoes as well as the quality of reconstructed speech.