Research And Implementation Of Echo Cancellation Algorithm In Smart Speaker

In recent years,with the continuous development of artificial intelligence technology,speech interaction has become the most convenient and efficient human-computer interaction mode,while smart speaker has gradually become the main human-computer interaction object.The quality of speech interaction is the key to determine the quality of human-computer interaction experience.Among them,the acoustic echo can seriously affect the quality of speech instructions to be recognized,thus the recognition of speech arousal and speech recognition results would be reduced.Therefore,echo cancellation in smart speaker environment has become a research hotspot in the field of artificial intelligence.The main research content is to solve the problem of inauthenticity and complexity of the traditional acoustic echo cancellation algorithm for the target speech signals,and the echo cancellation algorithm based on the smart speaker environment is proposed.The main research contents are as follows:First of all,based on the analysis and research of the basic theory and algorithm of echo cancellation,several traditional single-channel echo cancellation algorithms are elaborated and the evaluation criterion of echo cancellation performance is introduced.Secondly,considering that the complexity of the acoustic echo cancellation algorithm in smart speakers is proportional to the number of microphones,which can easily exceed the available computing resources.A multi-channel echo estimation algorithm of relative echo transfer function is proposed.The spatial structure of multiple channels is used to estimate the echo,which the computation would be reduced.On this basis,combining with generalized sidelobe canceller beamforming technology,a rapid echo cancellation algorithm for microphone array is proposed to obtain target speech signals with high source distortion and intelligibility.Thirdly,in order to solve the problem of poor robustness of echo cancellation in smart speakers,a joint echo and noise cancellation algorithm based on bidirectional long short-term memory neural network is proposed.The algorithm no longer requires the separate double talk detector,noise suppressor,and residual echo suppression module,so the problem of inaccurate is avoided effectively.The original speech signal information in frequency domain can be kept combing with removing noise and echoes,while speech intelligibility of the nonlinear echoes and non-stationary noise in the actual conditions is improved.Simulation results further verify the effectiveness of the algorithm.Finally,considering the microphone pickup performance for speech front-end algorithm(such as echo cancellation,the sound source localization,speech enhancement,etc.),the core function module including speech awakens and echoes also plays an important role,determining the smart speaker interaction quality.So,a platform for sound pickup performance of smart speakers based on microphone array is built,test indicators and test algorithms are proposed.The test results show that test conclusions can be achieved quickly,the accuracy of test results is improved,the difficulty of test is reduced on the platform.The better hardware support for echo cancellation algorithm and other subsequent algorithms is provided.