Research On Deep Learning Based Method For Audio Howling Processing And Suppression

In the enclosed room where the sound reinforcement system is used,the microphone and the loudspeaker are in the same environment,which leads to the existence of sound feedback,and in serious cases,whistling will occur.The harsh roar will affect the sense of hearing and damage the components of the sound reinforcement system,causing damage to the system.The existing methods of howling suppression include phase modulation(PM),gain control and adaptive acoustic feedback cancellation(AFC).According to the principle of phase modulation and gain control,the amplitude and phase of the signal will be destroyed.The adaptive acoustic feedback elimination method is limited by the performance of the adaptive algorithm.It has steady-state error and needs to add decorrelation technology to reduce the correlation between the signals.These methods will cause the degradation of sound quality.In order to solve the problem of howling,this paper proposes a deep learning method to study the problem of howling suppression.A supervised neural network is established to complete the mapping from the input howling signal to the output clean signal,so as to estimate and eliminate the howling feedback component in the input signal,achieve the purpose of Howling suppression,and minimize the distortion of the sound signal.In this paper,different technologies are used to study speech and audio signals,including input features,learning model(neural network structure),activation function and optimization algorithm.Through the experimental results,a more suitable combination is found to form the method proposed in this paper.Finally,the perceptual evaluation of speech quality(PESQ)and short-term objective intelligibility(STOI)of speech signal are improved,The SNR of audio signal is improved.Compared with the frequency-domain adaptive algorithm,the adaptive algorithm has the problems of convergence process and steady-state error,while the deep learning reduces the error of output signal and expected signal through the continuous learning of a large number of data,and outputs the prediction results directly during the prediction to eliminate the signal distortion in the convergence process.The comparative experiment results show that the improvement of speech quality and intelligibility after the howling signal is processed by the method in this paper is higher than that of the frequency domain adaptive algorithm.Moreover,from the time frequency diagram,the proposed method obviously eliminates the howling feature,which verifies the feasibility and effectiveness of the method.The method proposed in this paper has room to improve the sound quality.The following research can use the characteristics of human auditory characteristics,such as mel-frequency cepstral coefficient(MFCC),and the reconstruction of time-frequency signal will produce errors,the establishment of time-domain neural network is also a research direction.