1st Joint SIG-IL/Microsoft Workshop on Speech and Language Technologies for Iberian Languages
Porto Salvo, Portugal
In this paper a new algorithm is proposed for fast discriminative training of hidden Markov models (HMMs) based on minimum classification error (MCE). The algorithm is able to train acoustic models in a few iterations, thus overcoming the slow training speed typical of discriminative training methods based on gradient-descendent. The algorithm tries to cancel the gradient of the objective function in every iteration. Re-estimation expressions of the HMM parameters are derived. Experiments with triphone and word models show that the proposed algorithm always achieves much better results in a single iteration than MCE, MMI or MPE do over several iterations.
Index Terms: Speech recognition, discriminative training, hidden Markov models.
Bibliographic reference. Silva, B. / Mendes, H. / Lopes, C. / Veiga, A. / Perdigão, Fernando (2009): "A fast discriminative training algorithm for minimum classification error", In SLTECH-2009, 53-56.