Improving Automatic Speech Recognition with Dialect-Specific Language Models

We present an end-to-end Automatic Speech Recognition (ASR) system in the context of the recent challenge tasks for Bhojpuri and Bengali. Our implementation follows the currently popular wav2vec models while we investigate ways to leverage the dialect-categorised data in order to improve ASR performance. We report overall improvements in word error rate with dialect-specific language models for each of the languages. We present an analysis that provides insights into some of the factors underlying the success of dialect-specific language models.