This work describes a method for generating back-off models for context-dependent unit modeling. The main characteristic of the approach is that of building generic models by gathering statistics of detailed models, collected during Baum-Welch reestimation. The construction of back-off models does not require additional processing of the training data, allowing to quickly build different models sets with different back-off criteria starting from the same set of trained models and their statistics. Experiments are reported on the TIMIT and Wall Street Journal corpora, that show the consistency of the approach and compare it with state tying based on Phonetic Decision Trees.
Cite as: Brugnara, F. (2001) Model agglomeration for context-dependent acoustic modeling. Proc. 7th European Conference on Speech Communication and Technology (Eurospeech 2001), 1641-1644, doi: 10.21437/Eurospeech.2001-202
@inproceedings{brugnara01_eurospeech,
author={Fabio Brugnara},
title={{Model agglomeration for context-dependent acoustic modeling}},
year=2001,
booktitle={Proc. 7th European Conference on Speech Communication and Technology (Eurospeech 2001)},
pages={1641--1644},
doi={10.21437/Eurospeech.2001-202}
}