In this paper we proposed an efficient beam search procedure that combines well-known search techniques as a lexicon organization using tree-structured grammars with a novel approach of using different types of subword units depending on the local scores of the active words. An efficient double-tree structure using phonemes and triphones is presented. Experimental results on an isolated word recognition systems reveals that the proposed strategy is suitable for important reductions in computational cost with only negligible increases in recognition errors. Tests over a vocabulary of 955 Spanish words presents a 0.5% of increase in error rate for a 32% reduction in the number of senones to be evaluated.
Cite as: Torrecilla, J.C., Cortazar, I., Hernández, L.A. (1998) Double tree beam search using hierarchical subword units. Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998), paper 0321, doi: 10.21437/ICSLP.1998-670
@inproceedings{torrecilla98_icslp,
author={Juan Carlos Torrecilla and Ismael Cortazar and Luis A. Hernández},
title={{Double tree beam search using hierarchical subword units}},
year=1998,
booktitle={Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998)},
pages={paper 0321},
doi={10.21437/ICSLP.1998-670}
}