Research And Implement On The Extraction Of Speech Source Signal In Limited Space Based On OMAP-L138

In the buildings with limited space,mixing speech signals generated by speakers are composed of a pure speech source signal.The mixing speech signals are interfered by background noises and reverberations.As a consequence,the speech quality of the speech source signal degrades.The intelligibility and perceptibility decline.It is difficult to recognize.As a result,it is necessary to research on the issue of the extraction of a desired pure speech source signal within the mixing speech signals with noises.But,the algorithm research usually is a main research.The algorithm transplantation is short of research,especially using DSP with dual kernels.The blind source separation algorithm of the mixing speech signals with noises is researched,and the algorithm transplantation is researched based on OMAP-L138 with dual kernels.The main contributions of this paper are as follows:(1)In the blind source separation process of the mixing speech signals with noises,the algorithm in three phases including pre-processing denoising,blind source separation and post-processing dereverberation is researched respectively.Before the mixing speech signals with noises are separated,the signal to noise ratio needs to be improved necessarily and the distortion should be minimized possibly in order to ensure the accuracy of blind source separation.The improved power spectral subtraction algorithm based on the constant estimation of noises power spectrum is used to eliminate the noises.The simulation experiment has been done to verify that the proposed algorithm is suitable for the pre-denoising before blind source separation.When the denoised mixing signals are separated using blind source separation,the whitening of basic nature gradient algorithm is computed complexly.It is easy to bring errors.The iteration with fixed step is difficult to accelerate convergent speed and decrease steady-state errors.As a result,the improved nature gradient algorithm using self-adaptive step without the whitening is proposed.The simulation experiment has been done to verify that the performance is improved compared with basic nature gradient algorithm.After blind source separation,the main remains in speech source signal are reverberations.In order to suppress the remains fully,the algorithm using guided spectrogram filtering and wiener filtering is proposed.The simulation experiment has been done to verify that the proposed algorithm is better at suppressing remains and the speech source signal is purer.(2)In order to transplant and optimize algorithm on the OMAP-L138 development board,the hardware platform and software environment used to process the mixing speech signals with noises are set up.The process of algorithm transplantation has four phases,including developing,compiling,debugging and solidifying.At developing phase,the ARM used to develop human machine interface controls the DSP running.The DSP is used to compute algorithm.The ARM exchanges command and data with the DSP.At compiling phase,system project directory is built and compiled by Makefile file.At debugging phase,the system project is debugged dynamically on the ARM and DSP using GDB and CCS respectively.At solidifying phase,according to the executing method of the system project,the startup script file is made and set as a startup item for Linux.The result is got through the solidified system project executed.Finally,the analysis of the result is made through the contrast between transplantation and simulation.The analysis indicates that the transplanted algorithm has been processed well,and its overall processing time is reasonable.