SUBWORD UNITS FOR AUTOMATIC SPEECH RECOGNITION OF ANY VOCABULARY

Most commercially-available speech recognizers use whole-word modellin g techniques, with a model being trained for each word in the required vocabulary. A much more versatile system is obtained by using sub-wor d models, which offer the potential to build word models for any appli cation vocabulary from a single set of trained sub-word units. Ideally , a general database is used to train, once and for all a set of vocab ulary-independent sub-word models for the chosen language. In the past , systems of this type have tended to give considerably worse recognit ion performance than whole-word systems. The problem is to choose an a ppropriate sub-word unit, incorporating all important context effects within all possible words while not requiring an excessive number of w ords for training. To achieve this aim, a new type of sub-word unit (c alled the phonicle) has been developed at the Hirst Research Centre. T he paper provides some background by explaining the complexity of the speech signal and briefly introducing the most successful recognition techniques, before describing the phonicle approach in some detail. Sp eaker-dependent isolated-word recognition experiments on six example a pplication vocabularies (with an average vocabulary size of about 80 w ords) have demonstrated a very low average error-rate of 0.3%. The app roach is now being applied to speaker-independent and to connected-wor d recognition, for which it should have many commercial applications i n versatile voice-driven systems.