| In speech enhancement tasks,neural network based speech enhancement methods have strong nonlinear fitting capabilities,which can construct a mapping from noisy speech to clean speech.Utilizing the network does not require prior information such as the location of the sound source,which is different from some traditional array enhancement algorithms.Existing research work on speech enhancement networks usually only sets one or multiple objective functions at the output of the network.As the speech enhancement task becomes more complex,the network needs to complete more.This will lead to a sharp increase in the minimum points of the objective function.It is difficult to find a better solution if the network parameters are optimized only by back-propagation of the final output error.In this paper,we propose a multitask based microphone array speech enhancement method,which uses prior knowledge to subdivide a complex enhancement task into multiple enhancement tasks.The method can constrain and guide the optimization direction and process of network parameters,so as to reduce the probability of falling into the minimum value point,reduce the difficulty of training,and improve the effect of network enhancement.The work of this paper is as follows:(1)Aiming at the training difficulties and multi-target selection problems in the existing de-reverberation and noise reduction networks,this paper proposes a multi-task dereverberation and noise reduction network.The enhancement process of the multi-task network is divided into de-reverberation task,noise reduction task and multi-channel fusion task according to the functions of the speech enhancement system.Then we design corresponding subnet for each task and target loss function.The multi-channel fusion sub-network includes a speech channel convolution structure to better fuse the outputs of each channel branch network to enhance twice.The experimental results show that the proposed network has good generalization performance.Under different signal-to-noise ratios and reverberation strengths,the performance of the network is better than the existing traditional algorithms and methods based neural network.(2)In the remote voice communication scenario,the existing echo cancellation network methods have the shortcomings of echo residue and difficulty in training.We propose a multitask noise reduction system for remote communication.The network of the system divides the enhancement tasks into echo cancellation,de-reverberation,noise reduction and multi-channel fusion tasks.Each task has a corresponding sub-network.The input of the echo cancellation sub-network adopts the output of the existing linear echo cancellation algorithm,so as to reduce the training difficulty of the sub-network.In addition,in order to reduce the network training overhead,a multi-task noise reduction network is proposed which shares the parameters of multiple channel networks.The experimental results under dual-speaking and single-speaking conditions show that the two proposed networks outperform the comparative algorithms in performance. |
