In HMM-based pattern recognition, the structure of the HMM is often predetermined according to some prior knowledge. In the recognition process, we usually make our judgment based on the maximum likelihood of the HMM, without considering the time-varying property of state-level variables, which unfortunately may lead to incorrect results. In this paper, we analyze the property of state-level variables in the HMM and show it is possible to significantly enhance the performance of speech recognition systems when using the state-level variable time-varying property. We propose four methods to model state-level variable trajectories and then test them on a phoneme classification task on the TIMIT speech corpus, 11.95% error rate reduction is achieved and some empirical conclusions are drawn.
Cite as: Li, H.-Z., O'Shaughnessy, D. (2006) State-level variable modeling for phoneme classification. Proc. Interspeech 2006, paper 1332-Mon3BuP.5, doi: 10.21437/Interspeech.2006-220
@inproceedings{li06c_interspeech,
author={Hao-Zheng Li and Douglas O'Shaughnessy},
title={{State-level variable modeling for phoneme classification}},
year=2006,
booktitle={Proc. Interspeech 2006},
pages={paper 1332-Mon3BuP.5},
doi={10.21437/Interspeech.2006-220}
}