Research On Parallel Algorithm Of Deep Transductive Non-negative Matrix Factorization For Speech Separation

As a symbol of inheriting world civilization,voice signals have greatly enriched the development history of world civilization and have also become the focus of research in academia and industry.When voice signals are expressed in matrix form and calculated,they can solve problems in life.This paper focuses on speech separation as the research background,and studies the matrix decomposition effect and efficiency in the process of speech separation.First,due to the uncertainty of the separation scene,the multi-speaker mixed speech signal,there are problems such as insufficient signal feature expression,and the separation effect deviates from the original speech signal.Second,the speech separation process needs to update the iteration rule to make the matrix product approximate.Represents the original matrix.With the difference in the order of selection during matrix multiplication and the increase in the complexity of the problem,problems such as the increase in the amount of data in the calculation process,the low efficiency of the operation process,and the long operation time can not meet the real-time performance in the speech separation scenario Requirements,limiting the development of industrialization of speech separation.The research work is carried out on the existing problems in the speech separation scenario.The main work and results of this paper are as follows:1.In view of the problem of insufficient speech feature expression in the process of speech separation,this paper proposes an iterative and updated DTNMF optimization algorithm based on the deep transduction non-negative matrix factorization(DTNMF)algorithm.Based on a typical non-negative matrix factorization algorithm,the transductive learning idea is first introduced,and the mixed speech signal is creatively added to the objective function during the separation stage,so that the training data and the test data can be combined to generate a dictionary during the separation process;then Adding a deep structure,the base matrix is decomposed by layers during the iterative update process.Through each layer decomposition,the weight matrix of each layer can more accurately express the corresponding attributes until the pre-training process is completed.During the iterative decomposition process,set the threshold.When the convergence result of the objective function is greater than the threshold,the convergence speed is accelerated,otherwise it is slowed down.The experimental results show that the DTNMF method can effectively restore the speech signals of each speaker in the mixed speech signal,and obtain better results than the traditional NMF applied to the speech separation scene.2.Aiming at the problems of complex data calculation,too much calculation,and long calculation time in the separation process,a DTNMF parallel algorithm for speech separation is proposed.During the matrix multiplication operation,the matrix is divided according to rows and columns or summa algorithm blocks.Inter-process communication is used to exchange data,coordinate parallel computing processes,and accelerate the matrix computing process.The experimental results show that the MTN parallel algorithm based on DTNMF algorithm can effectively shorten the separation time and achieve a good acceleration effect in the separation of voice signal pre-training and separation process.3.Aiming at the possible load imbalance of the designed parallel algorithm,comprehensively consider the data correlation of the iterative update process,and design multi-level parallel algorithms between and within tasks.At the task level,the process of decomposing the training speech to obtain the corresponding base matrix is calculated in parallel as two independent tasks;at the internal process level of the task,the matrix is divided by rows and columns or by the summa algorithm block;during the thread-level submatrix multiplication,An acceleration strategy for generating multithreaded,sub-matrix block calculations through shared memory exchange data.This algorithm is the first parallel algorithm to implement deep transduction non-negative matrix factorization.The experimental results show that the proposed parallel algorithm can greatly reduce the time of separating the multi-speaker mixed speech signals without changing the separation effect,and can effectively improve the separation efficiency.