In this paper, a new system for the automatic segmentation and labeling of speech is presented. The system comprises segmentation and broad phonetic classification neural networks which were originally trained on one task (Flemish continuous speech), and which were subsequently adapted to a new task. The adaptation is performed by an embedded training procedure requiring no hand labeled utterances representative for the new task. The system was evaluated on five isolated word corpora designed for the development of Dutch, French, American English, Spanish and Korean text-to-speech systems. Additional test were run on TIMIT utterances in order to provide segmentation and labeling results which can be compared to similar results reported in the literature.
Cite as: Vorstermans, A., Martens, J.-P. (1994) Automatic labeling of speech synthesis corpora. Proc. 3rd International Conference on Spoken Language Processing (ICSLP 1994), 1747-1750, doi: 10.21437/ICSLP.1994-192
@inproceedings{vorstermans94_icslp,
author={Annemie Vorstermans and Jean-Pierre Martens},
title={{Automatic labeling of speech synthesis corpora}},
year=1994,
booktitle={Proc. 3rd International Conference on Spoken Language Processing (ICSLP 1994)},
pages={1747--1750},
doi={10.21437/ICSLP.1994-192}
}