Design And Implementation Of An Improved Dynamic Time Warping-based Speech Recognition Algorithm

With the rapid development of the science and technology,People started to pursue more intelligent and humanistic experience.Speech recognition technology has broad prospects in smart home applications.Household appliances can be controlled remotely with speech recognition technology.It can greatly improves the user experience.Speech recognition technology has advanced enough to be used on high-performance hardware platform.However,it still faces many challenges on embedded system.Therefore,speech recognition technology on embedded system has become research hotspots in recent years.The principle and basic model of speech recognition system are analyzed in this paper.The key techniques and difficulties on embedded system are analyzed.An improved endpoint detection and feature parameter extraction method is adopted.The match path between tested model and reference model is limited.After this,the amount of calculation and memory usage are reduced.Firstly,the threshold value of zero-crossing rate in endpoint detection is decreased to offset the side effect of noise.Secondly,an time period is added in the end point to ensure a valid result.Thirdly,the logarithm operation is removed and the order of parameter is elevated in feature parameter extraction process.The amount of calculation is reduced while the accuracy rate is not decreased.Finally,the match path of dynamic time warping algorithm is limited by rhomboid to avoid some unnecessary computation.The performance of the algorithm is improved and occupancy of system resources is decreased.Eventually,the algorithm is realized on embedded system.In this paper,an embedded speech recognition system is realized by SEP6210A.SEP6210A is a system on chip researched by ASIC of Southeast University.The function of the system is verified and compared in the related environment.The test results show that,the number of voice commands that the designed system can recognize is up to 45,the correct rate of speech recognition reaches 91%,the Real-time coefficient reaches 1.48.The results meet the design requirements.