A speech recognition and speech corpus system based on Matlab

In this paper, an educational mandarin syllable speech recognition and speech corpus system based on Matlab and SQL is introduced. The system includes a set of mixture CDHMM training and recognition algorithms both in Matlab M-files and optimized MEX-files, which can deal with multi-observations and scaling problems; a database toolbox to save and retrieve binary speech samples to and from a SQL database; a real-time speech recording ActiveX control to record speech samples and perform endpoint-detection. The three parts are integrated into Matlab to build up a speech recognition system mainly for educational purposes. It has been used in 408 mandarin syllables speech recognition; and a baseline recognition rate of 91.5% is reached.