New automatic speech recognition applications, mainly for small and medium vocabulary sizes, demand the capability of recognizing speech in several languages simultaneously. We have started exploring the possibility of building acoustic models that integrate multiple languages (up to seven in the initial stage), using speech transcriptions based on a common phoneme alphabet across all the languages. To reach a common alphabet, we start from the previously existing alphabets for each one of the seven languages. We first proceed to simplify some of them, partially following SAMPA transcription guidelines, and then to merge phones present in several languages that correspond to the same IPA symbol. We study and compare two variants of the common phoneme alphabet. The first of these two alphabets is closer to the starting ones, and includes the use of diphthong phones for English and German, and long-vowel phones for Arabic, English, and German. The second one avoids the long-vowel and diphthong phone models, and also the stressed vowel models. We present and discuss the results of decoding large vocabulary dictation tests, comparing the two alphabet variants, and also the multilingual decoding results with the corresponding monolingual acoustic models.
Cite as: Palou, F., Bravetti, P., Emam, O., Fischer, V., Janke, E. (2000) Towards a common phone alphabet for multilingual speech recognition. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 1, 501-504, doi: 10.21437/ICSLP.2000-124
@inproceedings{palou00_icslp,
author={Francisco Palou and P. Bravetti and O. Emam and V. Fischer and Eric Janke},
title={{Towards a common phone alphabet for multilingual speech recognition}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 1, 501-504},
doi={10.21437/ICSLP.2000-124}
}