Research On Speech Enhancement Of Small-Scale Microphone Array Based On Deep Learning

In recent years,speech has gradually become one of the most important methods for human-machine communication.In an enclosed room,noise and interference,as well as reverberation caused by reflections on walls and other surfaces,are the main reasons for the degradation of speech quality and intelligibility.Suppressing reverberation is a challenging task for speech enhancement due to high coherence between reverberation and speech.Noise and interference make the problem of indoor speech enhancement more difficult.Wind noise has non-stationary characteristics similar to speech signals,making outdoor speech enhancement more challenging.This thesis focuses on deep learning for indoor and outdoor speech enhancement.Further,array processing is combined with deep learning for enhancement of distant speech received by a multi-channel microphone data acquisition system.This thesis presents a deep neural network architecture for speech dereverberation based on combination of beamforming and deep complex U-Net(BF-DCUnet).Beamforming is utilized as a preprocessing module for improving signal-to-noise ratio(SNR)and suppressing interference from other directions.The dereverberation performance is further improved by integrating the frequency band extraction block and complex convolution operation.Meanwhile the heterogeneous convolution strategy is utilized to simplify the network model under a performance guarantee for dereverberation.On the base of THCHS-30 corpus,a reverberant speech data set is constructed for BF-DCUnet model training and simulated test.Finally,the processing results of the reverberant speech data collected by a multi-channel microphone data acquisition system have verified the effectiveness of the BF-DCUnet.For the task of speech noise reduction,in view of the problem that convolutional neural networks cannot effectively estimate the phase of desired speech signal,this thesis proposes a speech noise reduction algorithm based on deep complex convolutional neural networks and self-attention(DCCRN-SA),designing a complex convolution and dense block to achieve phase estimation.Proposed algorithm also suppresses non-stationary noise through combination of self-attention mechanism and complex Long-Short Term Memory(LSTM).Training the indoor and outdoor scene models based on Libri Speech corpus,and the DCCRN-SA performs well in tests using simulated data and noisy speech data collected by a multi-channel microphone data acquisition system.Finally,a speech enhancement system is designed and builted based on the Qt platform which includes two subsystems: one is for indoor speech enhancement with dereverberation and noise reduction;the other is for outdoor speech enhancement with wind noise suppression.The real-time processing results of the system have further validated the effectiveness of the proposed speech enhancement framework.