Utilization of Huge Written Text Corpora for Conversational Speech Recognition

Abstract

In this paper, we propose a new sentence selection method using large written text corpora to augment the language model of conversational speech recognition in order to resolve the insufficiency of in-domain training data coverage in conversational speech recognition. In the proposed method, the large written text corpora are clustered by an entropy-based method. Clusters similar to the target development set are selected automatically. Next, utterances are selected and mixed with the original conversational training corpus, and language models for conversational speech recognition are built. In our experiments, a different speech style test set that is not covered by original conversational training data is used for evaluation. The perplexity of the test set was reduced from 249.6 to 210.8, and the word recognition accuracy was improved by approximately 5% by using our method. Index Terms: data collection, training data coverage, language model, conversational speech recognition.