Speech Enhancement And Localization Based On Array And Neural Network

At present,Speech Enhancement(SE)and Sound Source Location(SSL)technology is frequently used in consumer electronic devices.SE and SSL have a close relationship.SE technology is used as the front end of SSL technology,which can improve the positioning performance of SSL and the performance of SSL also affects the performance of SE.However,the presence of ambient noise and reverberation can affect the quality and intelligibility of the speech signal and thereby reduce the performance of SSL systems.To solve these problems,microphone array technology and artificial neural network technology has emerged.Microphone array technology is developed to break through the limitations of single-microphone technology.Compared with traditional single microphone technology,it can accurately obtain a group of spatial information and suppress the noise as much as possible under the condition that the speech signal is not distorted.The combination of traditional array signal characteristics and artificial neural network technology can better solve the problem of traditional SSL,and it has been successfully applied in SSL field.On account of this,this paper studies SE and SSL technologies based on array and neural network.The main research contents are as follows:(1)The performance of Voice Activity Detection(VAD)algorithm is degraded under low signal-to-noise ratio.A scheme combining SE algorithm and VAD algorithm is proposed.The solution is to improve the signal-to-noise ratio of the original speech signal with the help of fixed differential beamforming and modulation domain spectral subtraction,and then use the improved spectral entropy algorithm for VAD detection.The experimental results show that the detection accuracy is mostly above 90% under low SNR,and the detection of the algorithm is better.(2)Aiming at a small double microphone speech denoising system,an Adaptive Noise Cancellation(ANC)algorithm based on VAD was applied to fixed difference beamforming with Log MMSE(Minimum Mean Square Error).At the same time,a time domain restoration method is proposed to recover distorted speech signals,which can obtain smaller time delay and computation than the existing frequency domain restoration algorithm.Theoretic analysis and experiments indicate that the proposed scheme can suppress directional noise effectively and easy for real-time implementation.(3)Aiming at the phenomenon of poor positioning accuracy of traditional algorithms,a Back Propagation(BP)neural network SSL algorithm based on correlation coefficient is proposed.The algorithm makes full use of the correlation between the received signals of microphones in the seven-element cross array to solve the correlation coefficient,takes the correlation coefficient as the input of BP neural network and the sound source position as the output.The simulation results show that the algorithm has better positioning performance.The average positioning error is 0.042 meters,and accuracy is 89.5%.Compared with the comparision algorithm,the error is reduced by 0.07 meters and the accuracy is improved by 18.5%.