Development of real-time embedded system with speech recognition for smart house

Abstract

The hidden Markov model is used for the acoustic modeling of the speech recognition system. The Continuous Density HMM (CDHMM) which models acoustic observation directly using estimated continuous probability density function (pdf) without VQ, has shown to have higher recognition accuracy than DHMM. In this paper, CDHMM with Multivariate Gaussian Density is used for speaker dependent (SD) and speaker independent (SI) Malay isolated digit recognition and comparison is made with DHMM. The CDHMM was trained by different algorithms- Baum-Welch (BW), Viterbi (VB) (with segmental k-mean estimation) algorithm and combination of BW and VB then comparison is discussed. The training database consisted of 26 speakers with 5 utterances for each Malay digit. In SD task, another 5 utterances each digit from the same speakers are used for testing. Recognition accuracy of CDHMM with BW, VB training and combination of BW and VB is 99.00%, 98.85% and 98.69% respectively while 96.62% for DHMM. The accuracy for BW and Segmental K-Mean training is comparable, but the latter consumed less computational time. In SI task, 40 speakers, different from the training speakers, with each recorded 2 tokens are used for testing. The CDHMM achieves 85.63% accuracy and outperform DHMM with 8.17% improvement.