A REAL-TIME CHINESE SPEECH RECOGNITION IC WITH DOUBLE MIXTURES

摘要

A Hidden Markov Model (HMM) based isolated word recognizer IC is presented in this paper. The test chip was fabricated with a 0.35 μ CMOS technology. In this design double-mixture probability density and table look-up method are employed. The chip can operate at 20MHz at 3.3V, and at this frequency the recognition time is 0.5 sec for a 50-word speech library. Tested with 353 speech data from AURORA 2 database, the chip's recognition accuracy is 93.8%. This new IC can be used in Cantonese speech recognition.