Unit Inventory and Model Structure for Vocabulary Independent HMM Recognition

This paper describes the use of allophonic sub-word units with an allophone-dependent model structure.to improve the performance of sub-word HMM recognition using vocabulary-independent training.The new system is an extension of an approach based on sub-triphone units called pltonicler.The original system I], 2] modelled major phonetic context effects by splitting each phone into a leftcontext dependent phonicle followed by aright-context dependent phonicle.It did not however take account of context effects wider than one immediately adjacent phone or the differences in duration and spectral complexity which exist between different types of phoneme.The recognition system has therefore been extended so that phoneme transcriptions are first converted to allophone transcriptions.Each allophone is then transformed to a sequence of one or more allophonicler.where different allophonicles can have different numbers of states and one allophonicle can be shared across allophones.Using a Mel Cepstrum front end.isolated-word speaker dependent recognition experiments on six example application vocabularies have shown a reduction in the average error rate from 4.9% to 0.3% by using allophonicle models.The paper discusses the results of this and other experiments in more detail.