5th International Conference on Spoken Language Processing
We choose to represent, unlike other teams, vocabulary words and out-vocabulary words with the same set of subword HMMs. Secondly we replace the classical one-phoneme transcription of fillers in the lexicon by a new, more powerful one-syllable transcription. As for the language model, the problem produced, in the case of unlimited-vocabulary continuous-speech recognition, by the lack of information on new words in the training corpus is solved through the use of the limited information we gathered on new words. The results obtained in general-task keyword spotting as well as unlimited-vocabulary continuous-speech recognition demonstrate the efficiency of the choice of a one-syllable transcription rather than a one-phoneme one. As for the results in unlimited-vocabulary continuous-speech recognition, the language model using information from words of frequency one is demonstrated to be a new promising method of determination of a language model for new words.
Bibliographic reference. Meliani, Rachida El / O'Shaughnessy, Douglas (1998): "Powerful syllabic fillers for general-task keyword-spotting and unlimited-vocabulary continuous-speech recognition", In ICSLP-1998, paper 0837.