Third International Conference on Spoken Language Processing (ICSLP 94)
This paper presents a contextual modeling of phonemes, and describes a new technique that renders a reliable estimation of contextual parameters. Using this approach the modeling of all of the acoustic realizations of a given sound is integrated into a single complex unit, for which each entry and exit state is assigned to a specific context. Context clustering trees are defined and used in order to provide a reliable estimation of the contextual parameters. Using hand-made trees, a 12 % reduction in the error rate is achieved on a 250-word vocabulary set, which is distinct from the training vocabulary. Finally, an automatic context clustering procedure is presented and applied in order to automatically generate these clustering trees. Using this automated approach the reduction in the error rate is comparable to that of the hand-made trees.
Bibliographic reference. Jouvet, D. / Bartkova, Katarina / Stouff, A. (1994): "Structure of allophonic models and reliable estimation of the contextual parameters", In ICSLP-1994, 283-286.