Parallel Optimization Technology Of Speech Feature Extraction Algorithms On Heterogeneous Multi-core Platform

Speech recognition is widely used in daily life.Speech feature extraction is one of the most crucial processes in speech recognition.While there is a problem of high time consuming and inefficiency,when speech Feature Extraction Faced with the current large-scale,highly computational intensive data computing problems.Traditional speech feature extraction algorithms can not meet their high-throughput requirements.Based on a variety of high performance computing platforms,this paper studied the parallel optimization techniques of three typical speech feature extraction algorithms:MFCC,SIF and Spectral Contrast.In this paper,for these three kinds of speech feature extraction algorithms,a CPU-based parallel algorithm is proposed to explore the impact of scheduling strategy on the performance of the algorithm.The experimental results show that the parallel optimization method based on the appropriate scheduling strategy can achieve greater performance improvement.In view of the powerful computational capability of GPU,this paper proposed a heterogeneous parallel algorithm for speech feature extraction based on GPU,and studied the effect of different task partitioning methods on the performance of the algorithm and a variety of parallel optimization techniques.It gets tens of times better speedup than the mainstream general-purpose microprocessor platform.Embedded devices have the advantage of high performance and low power consumption.Therefore,this paper studied a heterogeneous parallel algorithm of speech feature extraction based on embedded platform Jetson TX2.According to the characteristics of the platform,a Ping-pong cache structure is designed to further optimize algorithm,and it achieves 5.35 times speedup better than general-purpose microprocessor.