Boosting HMM acoustic models in large vocabulary speech recognition

摘要: Abstract Boosting algorithms have been successfully used to improve performance in a variety of classification tasks. Here, we suggest an approach apply popular boosting algorithm (called “AdaBoost.M2”) Hidden Markov Model based speech recognizers, at the level utterances. In recognition tasks show that significantly improves best test error rates obtained with standard maximum likelihood training. addition, results several isolated word decoding experiments may also provide further gains over discriminative training, when both training techniques are combined. our this holds comparing final classifiers similar number parameters and evaluating conditions lexical acoustic mismatch conditions. Moreover, present extension large vocabulary continuous recognition, allowing online without processing N-best lists or lattices. This is achieved by using for combining different models decoding. particular, introduce weighted summation extended set alternative pronunciation representing boosted baseline model. way, arbitrarily long utterances can be recognized ensemble single pass framework. Evaluation presented on two tasks: real-life spontaneous dictation task 60k Switchboard.