| 摘要 |
<p>The invention relates to a method of speech recognition based on a Hidden Markov model, whereby, during a first phase the known vocabulary is divided into a first partial vocabulary of important and/or interchangeable words and a second vocabulary of the remaining words. During a second phase the first partial vocabulary is trained with a whole word model and the second partial vocabulary is trained with a phoneme based model. During a third phase the transcriptions obtained from both models are stored in a single pronunciation lexicon.</p> |
