Boosted Large-Margin Estimation of hidden Markov models for speech recognition

We present a modified form of the Large Margin Estimation (LME) objective function which gives improved results for discriminative training. The modification consists of boosting the likelihoods of paths in the denominator lattice that have a higher phone error relative to the correct transcript, by using the same phone accuracy function that is used in Minimum Phone Error (MPE) training. The proposed Boosted LME method has also been successfully using the Microsoft corpora. The recognition results show that the proposed method yields better performance than the conventional approaches including maximum-likelihood estimation (MLE), maximum mutual information estimation (MMIE), and LME methods.