Context modeling and clustering in continuous speech recognition

Reports on the performance of two variants of well-known statistical-based clustering techniques and presents an evaluation on the TIMIT and TI-Digit databases. A clustering approach which (1) is based on a divergence criterion, (2) separates "good" and "bad" models using a class-dependent adjustable threshold on the number of examples per model, and (3) guides the clustering by limiting the number of models per class between two constants N/sub min/ and N/sub max/, gave the best results. On the TI-Digit database, the combination of triphone modeling and divergence-based clustering yielded greater accuracy than that obtained with word models for a similar system complexity.