Research On Multi-Sensor Speech Enhancement Algorithm For Wearable Devices

In recent years,with the rapid development of wireless voice communication technology,the quality of voice calls in quiet environments have been greatly improved.The sound field noise in the environment has become the major factor of interference with the speech quality and intelligibility.The microphone array technology is widely used as a voice enhancement technology that effectively suppresses noise.The microphone array is a kind of spatial filter,which can retain the desired speech signal and maximize the suppression of directional interference.In order to alleviate these problems,this thesis uses the advantages of microphone array processing technology to explore the use of a bone vibration sensor,which is not sensitive to air-conducted sound field noise on wearable device to improve noise suppression capability of the microphone array.The main work of this thesis is as follows:Firstly,based on the bone vibration signal is not interfered by complex noise environment.This thesis uses bone conduction speech presence probability to calculate robust voice activity detection.Then use the voice activity detection result to control the blocking matrix of the generalized sidelobe canceller and the update step size of the adaptive filter in the adaptive noise canceller.With the assistance of bone conduction signal,the filter coefficients of the blocking matrix and adaptive noise canceller can quickly converge in a complex sound field noise environment.Secondly,this thesis uses the characteristics of low correlation of wind noise between different microphones to detect the frequency bins interfered by wind noise,and designs a wind noise suppressor using the low-frequency harmonic components of the bone conduction signal.In order to further suppress the residual noise of the pre-array processing,this thesis propose a low-parameter real-time noise reduction neural network post-filtering algorithm based on simple recurrent unit.The post-filtering algorithm concatenate threeframe Mel energy spectrum as the model input,which reduces the redundancy of input features and improves the network's ability to suppress non-stationary noise.In addition,in order to alleviate the speech distortion,proposed algorithm uses the scale-invariant signal distortion ratio as the time domain loss function to assist the learning process of the network parameters in the training stage.Finally,experiments have proved that the desired speech leakage of the generalized sidelobe canceller blocking matrix can be effectively reduced,and the noise suppression capability of the adaptive noise canceller can be significantly improved and the algorithm improves speech quality and intelligibility compared to traditional methods.