Pronunciation dependent language models

Speech recognition systems are conventionally broken up into phonemic acoustic models, pronouncing dictionaries in terms of the phonemic units in the acoustic model and language models in terms of lexical units from the pronouncing dictionary. Here we explore a new method for incorporating pronunciation probabilities into recognition systems by moving them from the pronouncing lexicon into the language model. The advantages are that pronunciation dependencies across word boundaries can be modeled including contextual dependencies like geminates or consistency in pronunciation style throughout the utterance. The disadvantage is that the number of lexical items grows proportionaly to the number of pronunciation alternatives per word and that language models which could be trained using text, now need phonetically transcribed speech or equivalent training data. Here this problem is avoided by only considering the most frequent words and word clusters. Those new lexical items are given entries in the dictionary and the language model dependent on the chosen pronunciation. The consequence is that pronunciation probabilities are incorporated into the language model and removed form the dictionary, resulting in an error rate reduction. Also, the introduction of pronunciation dependent word pairs as lexical items changes the behavior of the language model to approximate higher order n-gram language models, also resulting in improved recognition accuracy.