Boosting Acoustic Models in Large Vocabulary Speech Recognition

C. Meyer and H. Schramm (Germany)


Boosting, AdaBoost, Machine Learning, Acoustic ModelTraining, Automatic Speech Recognition


The goal of this work is to evaluate the performance of boosting applied to acoustic model training in various tasks in large vocabulary automatic speech recognition. Specifi cally, we apply the AdaBoost.M2 algorithm -- at the level of utterances -- to maximum likelihood and discrimina tive training of the acoustic parameters of a Hidden Markov Model based speech recognizer. In an isolated word recog nition task, boosting improves the best test error rates ob tained with discriminative training, even when evaluating final classifiers with a comparable number of parameters. This is demonstrated in a matched and a mismatched de coding task. The second issue of our work is the extension of our algorithm to continuous speech recognition. To this end, we propose an approach realizing the combination of boosted acoustic models at a lexical level, allowing for an online (single-pass) decoding setup for the boosted models. First results are presented for maximum likelihood training in a real-life spontaneous speech dictation task with a 60k word vocabulary and about 58h of training data.

Important Links:

Go Back