This paper describes an improved phoneme-history-dependent (PHD) search algorithm. This method is an optimum algorithm under the assumption that the starting time of a word depends on only a few preceding phonemes (phoneme history). The computational cost and number of recognition errors made by a multi-pass-based recognizer can be reduced if the PHD search of the first decoding pass uses re-selection of the preceding word and the optimum length of phoneme histories. These improvements increase the speed of the first decoding pass and help that the word lattice has the correct word sequence. Consequently search errors can be reduced in the second decoding pass. In 65k-word domain-independent Japanese read-speech dictation task and 1000-word spontaneous-speech airplane-reservation task, the improved PHD search was 1.2-2.0 times faster than a traditional word-dependent search under the condition of equal word accuracy.
Cite as: Hori, T., Noda, Y., Matsunaga, S. (2001) Improved phoneme-history-dependent search for large-vocabulary continuous-speech recognition. Proc. 7th European Conference on Speech Communication and Technology (Eurospeech 2001), 1809-1813, doi: 10.21437/Eurospeech.2001-427
@inproceedings{hori01b_eurospeech,
author={Takaaki Hori and Yoshiaki Noda and Shoichi Matsunaga},
title={{Improved phoneme-history-dependent search for large-vocabulary continuous-speech recognition}},
year=2001,
booktitle={Proc. 7th European Conference on Speech Communication and Technology (Eurospeech 2001)},
pages={1809--1813},
doi={10.21437/Eurospeech.2001-427}
}