A word like phonetic sequence statistical source model for large lexicon automatic speech recognition systems

Here is presented a phonetic source model whose parameters, estimated from phonetically transcribed texts, reflect the non-stationary phoneme conditional probability which is proper of a given language. Such a model will give a priori knowledges about the allowed phonetic sequences probabilities for a very large vocabulary speech recognizer, where the lexical access is made after phonetic decoding. After a discussion about the probability estimation method, model features and performances are given.