Study On Speech Denoising And Post-Processing Technology For Wearable Devices

In recent years,with the rapid development of artificial intelligence and mobile communication technology,the research of wearable devices has become a hot field.Wearable devices mainly include wireless Bluetooth headsets,bracelets and other types.People can achieve human-computer interaction and human-to-human communication through wearable devices.Usually,there is a lot of noise,human voice and other interference in the sound field environment where the wearer is located,which will seriously affect the wearer’s voice experience.Most wireless Bluetooth headsets use microphone array voice enhancement technology to ensure the quality of the voice calls in complex sound field environments.The generalized sidelobe canceller is one of the classic algorithms in the microphone array beamforming algorithm.This thesis mainly investigates an improved algorithm for a robust generalized sidelobe canceller which is assisted by speech activity detection trained by a recurrent neural network.It can be applied to voice enhancement in wireless Bluetooth headset scenarios to improve the voice quality and intelligibility of human-computer interaction and interpersonal calls.At the same time,in order to further eliminate the residual noise in the output signal of the robust generalized sidelobe canceller,an improved post-processing technique is also proposed in this thesis.The main contents of this thesis are as follows:First,the blocking matrix module in the traditional generalized sidelobe canceller has the problem of expected speech signal leakage,which will lead to unsatisfactory speech quality after algorithm processing.In response to this problem,this thesis proposes a method to control the update of adaptive blocking matrix filter coefficients using speech activity detection trained by recurrent neural networks.Considering that wireless Bluetooth headset usually uses dual microphone linear array to receive voice signals,this thesis improves the input mode of cyclic neural network to dual channel voice data input,and sets the upper branch of robust generalized sidelobe canceller as dual microphone input.The experimental results show that the improved algorithm can effectively improve the performance of the adaptive blocking matrix,reduce the leakage of the desired speech signal,and prevent the reference noise from being eliminated by mistake.Thereby,it can reduce the speech distortion in the wireless Bluetooth headset application.Secondly,the generalized sidelobe canceller is mainly used for coherent noise,point source noise,stationary noise and other noise types.However,incoherent noise,weakly coherent noise and musical noise still remain in the speech signal after processing.In order to further eliminate the residual noise in the output signal of the robust generalized sidelobe canceller to ensure the quality of voice calls in the wireless Bluetooth headset scenario,this thesis proposes an improved post-processing technique with a combined ratio parameter of a recurrent neural network voice activity detector.Compared with the traditional postprocessing algorithm,the improved algorithm requires less computation.At the same time,it can estimate the noise spectrum more accurately,and can effectively suppress the incoherent noise and musical noise in the output signal of the generalized sidelobe canceller.The experimental results demonstrate the effectiveness and robustness of the proposed improved algorithm.