Implementation and Optimization of Embedded Speech Recognition System Based on DHMM

In order to meet the demand of realizing the speaker-independent speech recognition on the embedded system, a embedded speech recognition system in discrete hidden Markov model (DHMM) is proposed. The software architecture of the recognition software part running on embedded system and the training software part running on PC system are designed. According to features of TMS320VC5509A processor, the recognition software on embedded system is implemented and optimized. Also it is presented that the Viterbi algorithm is optimized on the basis of the irreversible characteristics of the speech signal. Experiments show the real-time response speed of the system achieved within 200 ms at same the recognition rate, and show its high availability.