The paper presents a software architecture allowing to collect, select, and exploit speech data from a specific application field to dynamically generate Hidden Markov Models tailored to that application environment and vocabulary. The framework we are interested in is, therefore, an already operational voice activated service that allows to collect directly from the field a large amount of speech data. We propose a procedure for data selection and for incremental training of the units using a strategy of model selection. Several tests are presented for a train timetable information system, and for a Directory Assistance application with a very large vocabulary of city names showing that significant improvements can be obtained with respect to the laboratory models, keeping the old models and transcribing only the most frequent words in terms of the new units, incrementally trained from the field data.
Cite as: Vair, C., Fissore, L., Laface, P. (2000) Dynamic adaptation of vocabulary independent HMMs to an application environment. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 2, 839-842, doi: 10.21437/ICSLP.2000-400
@inproceedings{vair00_icslp,
author={Claudio Vair and Luciano Fissore and Pietro Laface},
title={{Dynamic adaptation of vocabulary independent HMMs to an application environment}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 2, 839-842},
doi={10.21437/ICSLP.2000-400}
}