Large scale discriminative training for speech recognition

摘要: This paper describes, and evaluates on a large scale, the lattice based framework for discriminative training of vocabulary speech recognition systems Gaussian mixture hidden Markov models (HMMs). The concentrates maximum mutual information estimation (MMIE) criterion which has been used to train HMM conversational telephone transcription using up 265 hours data. These experiments represent largest-scale application techniques authors are aware, have led significant reductions in word error rate both triphone quinphone HMMs compared our best trained likelihood estimation. MMIE latticebased implementation used; ensuring improved generalisation; interactions with adaptation all discussed. Furthermore several variations scheme introduced aim reducing over-training.