An approach to efficient generation of high-accuracy and compact error-corrective models for speech recognition

This paper focuses on an error-corrective method through reranking of hypotheses in speech recognition. Some recent work investigated corrective models that can be used to rescore hypotheses so that a hypothesis with a smaller error rate has a higher score. Discriminative training such as perceptron algorithm can be used to estimate such corrective models. In discriminative training, how to choose competitors is an important factor because the model parameters are estimated from the difference between the reference (or oracle hypothesis) and the competitors. In this paper, we investigate the way how to choose effective competitors for training corrective models. Particularly we focus on word error rate (WER) of each hypothesis and show that a higher WER hypothesis rather than the bestscored one works effectively as a competitor. In addition, we show that using only one competitor with the highest WER in an N-best list is very effective to generate accurate and compact corrective models in experiments with the Corpus of Spontaneous Japanese (CSJ).