Very large vocabulary isolated utterance re comparison between one pass and tw

摘要: A system for recognizing isolated utterances belonging to a very large vocabulary is presented that follows two pass rategy. The first step, hypothesieation, consists in the lection of subset word candidates starting from gmentation speech into 6 broad phonetic classes. This odule implemented through Dynamic Programming gorithm working three dimensional space. search performed on tree representing coarse description e lexicon. second step best N according maximum likelihood criterion. Each candidate represented by graph sub-word Hidden Markov Models and structure whole built line an efficient implementation rison done with direct ape hypothesization module apof approach, which has same lice 80% reduction computational com