Utterance-level boosting of HMM speech recognizers

We propose an utterance-level approach to boosting HMM speech recognizers. The standard AdaBoost.M2 algorithm is applied to calculate training weights for utterances, which are used in maximum likelihood (ML) training of subsequent acoustic models. In recognition, scores of these models are linearly combined. We evaluate our algorithm on a large vocabulary isolated word recognition task. Significant performance improvements (up to 9% relative) are obtained compared to the ML baseline model, even for high density acoustic models. In particular, combining scores from boosted models outperforms combining scores from different ML baseline models.