INTERSPEECH 2008

We perform large margin training of HMM acoustic parameters by maximizing a penalty function which combines two terms. The first term is a scale which gets multiplied with the Hamming distance between HMM state sequences to form a multilabel (or sequence) margin. The second term arises from constraints on the training data that the joint loglikelihoods of acoustic and correct word sequences exceed the joint loglikelihoods of acoustic and incorrect word sequences by at least the multilabel margin between the corresponding Viterbi state sequences. Using the softmax trick, we collapse these constraints into a boosted MMIlike term. The resulting objective function can be efficiently maximized using extended BaumWelch updates. Experimental results on multiple LVCSR tasks show a good correlation between the objective function and the word error rate.
Bibliographic reference. Saon, George / Povey, Daniel (2008): "Penalty function maximization for large margin HMM training", In INTERSPEECH2008, 920923.