EUROSPEECH 2001 Scandinavia
This paper presents an approach to the integratation of contextual phonological rules in the beam-search algorithm of a large vocabulary speech recognition system. The main interest of contextual transcription rules is that they implement constraints on pronunciations sequences which complement the bigram constraints on word sequences. As such, they should help avoiding acoustic confusions and reduce the search space. In our approach, contextual transcription do not incur any augmentation of the lexicon size. This approach is evaluated on a dictation task in French for two different sets of contextual phonological rules. Our results show that, given the current resources, the introduction of contextual rule deteriorates the recognition rate. We discuss the possible factors explaining this surprising result and outline the problems of defining a set of contextual phonological rules and integrating them in the search algorithm.
Bibliographic reference. Gravier, Guillaume / Yvon, Francois / Jacob, Bruno / Bimbot, Frederic (2001): "Integrating contextual phonological rules in a large vocabulary decoder", In EUROSPEECH-2001, 2293-2296.