Combining multiple end-to-end speech recognition models based on density ratio approach

Automatic speech recognition (ASR) systems which involve the use of multiple ASR models, known as "system combination" approaches, have been confirmed by many studies to be effective for improving speech recognition performance. In this study, we propose a system combination method using multiple end-to-end (E2E) ASR models trained under different conditions (i.e., using data from different source domains), in order to perform ASR on an unknown target domain without the need for additional, expensive model training. In order to efficiently adapt existing ASR models to the target domain, we exploit a density ratio approach (DRA), and experimentally demonstrate that by combining ASRs trained with multiple types of acoustic information with an external language model for the target domain, i.e., the standard DRA, our method can more robustly perform ASR for an unknown target domain than a single ASR model trained under one domain condition.