Research On Multi-point Real-time Voice Transmission In Lan Based On Mobile Terminal

With the development of communication technology,a series of communication software,such as WeChat,QQ,Fetion,etc.,have been developed,which has brought great convenience to our daily communication.However,the consumption of a large amount of traffic leads to higher costs in voice communication or video communication.The emergence of WiFi solves this problem.People can access the network anytime and anywhere at a lower cost with no limitation by the duration.WiFi signals are popular because of their low price and convenience.Nowadays,many enterprises and companies are more inclined to choose communication software that can be used internally and whose communication environment is relatively safe.Moreover,software whose network cost is relatively small will also be considered.Therefore,the development of such application software will have a broader market prospect.At present,the research on multi-user communication under the local area network is very much.This paper will focus on the realization and approach optimization of voice communication.There are many attempts on real-time voice calls based on mobile terminals in the local area network,and significant results have been achieved.However,the problems of noise,mixing audio and echo in the communication process have not been well solved.This article will combine the deep learning algorithms to try to solve the above difficulties and improve the call quality.Firstly,a speech denoising approach based on deep learning for the noise problem existing during the call is introduced.A speech enhancement algorithm based on joint convolution-recurrent neural network is proposed.This algorithm is driven by data,which can accurately model the sound signal and capture the signal characteristics.However,the algorithm has a deficiency: when the sound sequence is too long,the information carried by the signal at the far end of the sequence cannot be accurately captured,thereby causing interference to the prediction and affecting the overall performance of the model.In order to solve this problem,a convolution-recurrent neural network algorithm based on self-attention mechanism is proposed,which overcomes the problem of long-term dependence of signal sequences.In this frame,even if the signal at the far end of the sequence can be captured.So the extracted sound feature vector is expressed more accurately and clearly.Secondly,in order to solve the problem of accurate mixing audio in multi-user conversation,an adaptive weighting-based mixing scheme is presented,which introduces a variable attenuation factor.It can make adjustments with the change of the sound so that the audio can change smoothly and the distortion of the audio can be reduced.The method can well solve the problems of signal overflow,waveform distortion,popping sound,etc.,which are caused by traditional mixing algorithms such as linear weighting and clamp method.It can effectively improve the call quality.Finally,an echo cancellation scheme based on improved equalization normalization for theecho problem introduced in the call is proposed.This approach assigns the corresponding gain to the coefficients of each adaptive filter according to a certain ratio,which solves the problem that the traditional minimum mean square algorithm has poor adaptability to non-stationary signals.It can accelerate the convergence speed of the adaptive filter when estimating the echo path,improving the efficiency of the algorithm operation and weakening the echo in the call.