Research And Implementation Of Real-Time Echo Cancellation Algorithm In Conference Telephone

With the rapid development of the Internet industry,various new types of communica-tion devices have emerged in an endless stream,greatly facilitating people's lives.With its high quality,low cost and convenient advantages,VoIP conference phones have quickly gained popularity among a large number of users.In the conference call,the acoustic coupling between the microphone and the speaker makes the echo phenomenon more serious,which greatly reduces the user experience.To ensure the communication effect,the echo must be effectively eliminated.Based on the in-depth study of the classical echo cancellation algorithm,this paper combines the popular deep learning technology to try to apply the deep learning network to echo cancellation.The main research contents of this paper are as follows:(1)The research background and significance of echo cancellation are introduced,and the research history and current situation of echo cancellation related technologies at home and abroad are summarized from theoretical research and system hardware.(2)The structure of the classical echo cancellation algorithm is introduced.First,the basic knowledge of acoustic echo cancellation is introduced,including the causes of acoustic echo,the working mode of conference calls and the preprocessing of speech signals.Then the structure of the commonly used classical echo cancellation algorithm is introduced in detail.The key modules of linear adaptive filtering,nonlinear post filtering and double-talk-detection are introduced.These classic algorithms provide architectural support for the algorithms presented in subsequent chapters of this paper.(3)The commonly used deep learning algorithms and their applications in the field of speech separation are introduced.Firstly,the calculation of the common feature parameters of speech(such as short-term average amplitude,short-time zero-crossing rate,pitch period,MelFrequency Cepstral Coefficients,etc.)is introduced.Then the basic theories of pattern recognition,machine learning and deep learning,and the connections and differences between them are explained.Then several commonly used deep learning algorithms are introduced.Finally,the application of the current deep learning technology in the field of speech separation is introduced.These applications are the inspiration for the subsequent research in this paper.(4)An echo cancellation structure based on DNNs is proposed.The structure uses the DNNs neural network to replace the adaptive filter in the classical algorithm to characterize the echo path.By training the DNNs neural network,the mapping relationship between the reference signal and the echo signal is learned,so that the echo can be mapped according to the reference signal to be eliminated.This chapter first introduces the principle of each module of the structure,then introduces the experimental database,the training and test dataset production methods.Finally,the effectiveness of the structure is verified by simulation.(5)An echo cancellation structure based on LSTM-RNNs is proposed.The structure can be divided into two parts,one part is used to learn the mapping relationship between the reference signal feature and the echo signal feature,and the other part is used to learn the mapping relationship between the desired gain,the reference signal feature and the near-end signal feature.This chapter first introduces the Recurrent Neural Networks(RNNs)that perform well in processing time series data,and introduces two improved versions of the RNNs structure LSTM and GRU proposed to solve the long-term dependency problem.Then the details of each module principle of the structure proposed in this chapter are introduced in detail.Finally,the effectiveness of the structure is verified by simulation.(6)Summarized the work done,focusing on the study of classical algorithms and the two architectures that use deep learning networks for echo cancellation based on the application of deep learning in de-reverberation and speech enhancement.Finally,the future work is prospected from the generalization and practicability of the algorithm.