This paper proposes a novel architecture for language-independent ASR based on government phonology (GP). We use experimental data to show that phoneme-based recognisers perform poorly on languages other than the original target, rendering such systems inadequate for multi-lingual speech recognition, a result we attribute to the inadequacy of the phoneme as a linguistic unit. In the proposed GP model, recognition targets are a small set of sub-segmental primes, or "elements", found in all languages, which have been previously shown to be robustly detected in a language-independent manner. Well-formedness constraints are captured by simple parameter settings which can be easily encoded as rules and applied as top-down constraints in a speech recogniser. Hence, given a set of trained element detectors, a recogniser for any given language can in principle be rapidly built by selection of the appropriate lexicon and constraints. We describe the design of experimental architectures for our GP-based system.
Cite as: Williams, G., Terry, M., Kaye, J. (1998) Phonological elements as a basis for language-independent ASR. Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998), paper 0622, doi: 10.21437/ICSLP.1998-774
@inproceedings{williams98c_icslp,
author={Geoff Williams and Mark Terry and Jonathan Kaye},
title={{Phonological elements as a basis for language-independent ASR}},
year=1998,
booktitle={Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998)},
pages={paper 0622},
doi={10.21437/ICSLP.1998-774}
}