A new framed-based MCE training strategy for Mandarin LVCSR

Minimum classification error training (MCE), as a method of discriminative training, has proved successful to reduce error rates in HMM-based speech recognition systems. Several drawbacks of the classical MCE training method, such as high computation complexity, small amount of contributing speech segments, and the difficulty to determine the learning rate, have been improved by a new training framework proposed in this paper. The new framework includes (i) predefining the word lattice for each utterance to limit the search space; (ii) selecting frame-based N-top active models as competing model sets of the correct model; (iii) defining different learning rates for different model parameters experimentally. We give experimental results on the 863 Mandarin speech corpus. Our method gives a 2.0% increase of the word-level accuracy rate of syllable loop-back search compared to MLE trained models by two iterations