Phonemic classification is still one of the most popular approach to automatic speech recognition. We address the problem of phonemic classification performance in a multi speakers and speaker independent environment A new tecnique for pattern classification is introduced The method is based on partial reconstruction of the target vectors, representing the speech signal in different parametrization forms, using projections on privileged orthonormal bases. The significative different results, obtained in function of the training and test experiments, in case of multi speakers and speaker independent evaluation, are discussed. Italian digits, that phonemically cover about half of the Italian phonemic vocaburary, was used as training and test signal. A total of 250 sentences spoken by five male speakers was used in this preliminary experiment. Phonemic classification performance moves around 50% for the speaker independent experiment, while in a multi speakers environment reachs the 65% of correct classification. The proposed method seems to be, although in its preliminary idiot-proof form, a good pattern matching solution for different speech parametrization in spite of its simple theorical justification, and its very low computational complexity.
Cite as: Falcone, M., Paoloni, A. (1990) Phonemic classification in mufti speakers and speaker independent environment. Proc. ESCA Workshop on Speaker Characterization in Speech Technology, 98-105
@inproceedings{falcone90_scst,
author={Mauro Falcone and Andrea Paoloni},
title={{Phonemic classification in mufti speakers and speaker independent environment}},
year=1990,
booktitle={Proc. ESCA Workshop on Speaker Characterization in Speech Technology},
pages={98--105}
}