Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

MMIE Training for Large Vocabulary Continuous Speech Recognition

Yves Normandin, Roxane Lacouture, Regis Cardin

Centre de recherche informatique de Montreal (CRIM), McGill College, Montreal Quebec, Canada

Over the past few years, there have been several published reports on the use of MMIE (Maximum Mutual Information Estimation) for training HMM parameters. Many of these reports clearly demonstrate MMIE's capability of substantially reducing the error rate in certain types speech recognition applications. The most convincing demonstrations, however, have usually been carried out on either small vocabulary tasks or on isolated speech. This is in large part a consequence of the fact that, for large vocabularies, MMIE training is much too computationally expensive and approximations must therefore be found in order to bring the task down to a manageable size. This paper looks at such approximations which can be applied in the context of large vocabulary continuous speech recognition and, in particular, proposes a technique based on the use of a compact looped word lattice. Experiments are used to demonstrate the effectiveness of the technique.

Full Paper

Bibliographic reference.  Normandin, Yves / Lacouture, Roxane / Cardin, Regis (1994): "MMIE training for large vocabulary continuous speech recognition", In ICSLP-1994, 1367-1370.