In this paper, selective strategy about MCE based discriminative training method,in particular for mandarin syllable loop recognition,is introduced. The basic idea is that since the decoding errors occur in parts of the models in whole decoded sentence, it is reasonable to adjust the parameters of the "wrong models". As a result, weighted MCE formulation is derived, which can provide more effective convergence property and about 10% error rate reduction for a large training set is achieved. On the other hand, from our experiments, we observed that although the whole performance of recognition system is improved, some original correct recognition results are misrecognized after discriminative training, divide and conquer strategy is proposed to solve it. Combining above two methods, we got more than 14.5% error reduction in syllable loop recognition experiments.

Full Paper

Bibliographic reference.  Zhou, Jianlai / Chang, Eric / Huang, Chao (2001): "Selective MCE training strategy in Mandarin speech recognition", In EUROSPEECH-2001, 1951-1954.