METHOD FOR RECOGNIZING SPEECH USING LINGUISTICALLY-MOTIVATED HIDDEN MARKOV MODELS

摘要

The invention relates to a language model for words using a state transition diagram (10). This diagram connects states (12, 14, 16, 18, 20) to arcs (22, 24, 26, 28, 30, 32) which represent transitions into phonemes within a word. Different pronunciations of the same word can be put in model using alternative transitions (26 or 28 and 30 or 32) from simple states (16 and 18 respectively). Is thus obtained alternative pronunciations models using similar phonemes, such as, in the word 'water' '(10), by substituting a phoneme' 'D' '(28) the phoneme' 'T' '(26 ). These models can be extended to the treatment of longer words or connected to each other for the treatment of several words.