Abstract: Non-negative matrix factorization can preserve the non-negative features of the speech signal.It is an important method for speech separation.However, this method has the problems of complicated data operation and computation, it is necessary to propose a parallel method to reduce computation time.Aiming at the calculation problem of speech separation pre-training and separation process, this paper proposes a deep transductive non-negative matrix factorization multi-level parallel algorithm, which considers the data correlation of the iterative update process, and designs a multi-level parallel algorithm between tasks and within tasks.The parallel algorithm at the task level decomposes the training speech to obtain the corresponding base matrix as two independent tasks in parallel calculation.The matrix is divided into rows and columns at the internal process level of the task.The master process distributes the matrix blocks to the slave process, and the slave process receives the current sub-matrix, then the matrix block calculates the result matrix sub-block, and then sends the current process matrix block to the next process to achieve the traversal of each matrix block of the second matrix in all processes, calculating the product of the corresponding sub-block of the result matrix, and finally the sub-block is collected by the main process from the slave process.During the thread-level sub-matrix multiplication operation, an acceleration strategy of generating multiple threads and exchanging data through shared me-mory for sub-matrix block calculation is adopted.This algorithm is the first one to implement deep transduction non-negative matrix factorization algorithm.The experimentis performed on the Tianhe II platform.The test results show that when separating multi-speaker mixed speech signals without changing the separation effect, in the pre-training process, the proposed parallel algorithm performs well using 64 processes at a speed ratio of 18, and in the separation process, the corresponding speedup is 24.Compared to serial and MPI model separation, hybrid model separation time is greatly shortened, which proves that for the speech separation process, the parallel algorithm proposed in this paper can effectively improve the separation efficiency.

Key words: Acceleration algorithm based on multiplicative update rules, MPI, OpenMP, Parallel algorithm of deep-transductive non-negative matrix factorization, Speech separation