Efficient Discriminative Training of Error Corrective Models Using High-WER Competitors

摘要

We focus on error corrective models for speech recognition, which select a more accurate word sequence in a word N-best list produced by a speech recognizer. In general, an error corrective model is trained so that the correct word sequence is discriminated by the model from all the sequences in each N-best list. However, we show that a more accurate model can be provided through the discrimination of only two word sequences; one is the correct word sequence and the other is the sequence with the highest word error rate (WER) in the list. We think there are two reasons for this effect: (1) many important error patterns can be obtained since typical errors frequently appear in the word sequences with high WERs, and (2) training with fewer word sequences alleviates the difficulty of parameter estimation. In addition, the model size becomes more compact and the training time is significantly reduced with our method. Experiments using the Corpus of Spontaneous Japanese (CSJ) show that our proposed method generates accurate and compact models, and specifically performs robustly for different tasks and linguistic features.