Parallelizing Research And Implementation Of Speech Recognition

The study object of speech recognition is voice; the study purpose of speech recognition is to enable machines to recognize human’s voice; the study content of speech recognition is how to make machines to extract speech information from speech signal in all environments, and machines can effectively understand and implement all kinds of intentions of speakers.Speech recognition system has two processes:training process and recognition process. When speech recognition system is implemented in serial mode, the running time of training process is too long; and the real-time performence of recognition process is sometimes difficult to meet the need of application environments. Parallel processing can solve these problems, therefore, the thesis will research on how to parallel the speech recognition system.The current mainstream processor is the multi-core processor with parallel data processing unit; therefore, multi-core processor can both meet the hardware requirements of multi-thread parallel and data processing parallel. The thesis will use OpenMP and SSE instructions to implement speech recognition system, and which can reduce the execution time and improve the real-time performance. The thesis implements four versions of isolated word speech recognition system:the version which implements in serial mode, the multi-thread version which implements in OpenMP, the data parallel version which uses SSE instructions, and the parallel version which both uses OpenMP and SSE instructions. Finally, the thesis compares the performance of the different parallel versions of speech recognition system with the serial version, and gets a good parallel performance.