Alignment of speech with a phonetic representation using continuous density hidden Markov models

Sentence models are constructed from 7-state hidden Markov models utilising tied transitions within the Markov model and tied states within the sentence model. The HMMs generate output on state-to-state transition. The training is performed using unmarked sections of speech of which only the phonetic content is known. This paper also gives the formulae used for the training including the tied transition and null transition instances, as derived by the authors from the Baum re-estimation training algorithm. Scaling of probabilities is also discussed