Errgrams – A Way to Improving ASR for Highly Inflected Dravidian Languages

In this paper, we present results of our experiments with ASR for a highly inflected Dravidian language, Telugu. First, we propose a new metric for evaluating ASR performance for inflectional languages (Inflectional Word Error Rate IWER) which takes into account whether the incorrectly recognized word corresponds to the same lexicon lemma or not. We also present results achieved by applying a novel method – errgrams – to ASR lattice. With respect to confidence scores, the method tries to learn typical error patterns, which are then used for lattice correction, and applied just before standard lattice rescoring. Our confidence measures are based on word posteriors and were improved by applying antimodels trained on anti-examples generated by the standard N-gram language model. For Telugu language, we decreased the WER from 45.2% to 40.4% (by 4.8% absolute), and the IWER from 41.6% to 39.5% (2.1 % absolute), with respect to the baseline performance. All improvements are statistically significant using all three standard NIST significance tests for ASR.