Allophone modeling for vocabulary-independent HMM recognition

The authors describe the use of subword units based on allophones with an allophone-dependent model structure, to improve subword HMM (hidden Markov model) recognition performance when using vocabulary-independent training. The new system is an extension of an approach based on sub-triphone units called phonicles. The original system modeled major phonetic context effects, but did not take account of context effects wider than one immediately adjacent phone or the differences in duration and spectral complexity which exists 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 allophonicles, where different allophonicles can have different numbers of states and one allophonicle may be shared across allophones. Using a Mel cepstrum front end, isolated-word speaker-dependent recognition experiments on six application vocabularies have shown extremely good recognition performance for allophonicle models, with an average error rate of only 0.3%.<>