Transfer learning of language-independent end-to-end ASR with language model fusion
This work addresses the challenge of building effective ASR for low-resource languages, though it appears incremental as it builds on existing transfer learning and LM fusion methods.
The researchers tackled the problem of adapting automatic speech recognition (ASR) systems to low-resource languages by using transfer learning with language model fusion, resulting in improved performance across five target languages and significantly reducing the gap with hybrid systems.
This work explores better adaptation methods to low-resource languages using an external language model (LM) under the framework of transfer learning. We first build a language-independent ASR system in a unified sequence-to-sequence (S2S) architecture with a shared vocabulary among all languages. During adaptation, we perform LM fusion transfer, where an external LM is integrated into the decoder network of the attention-based S2S model in the whole adaptation stage, to effectively incorporate linguistic context of the target language. We also investigate various seed models for transfer learning. Experimental evaluations using the IARPA BABEL data set show that LM fusion transfer improves performances on all target five languages compared with simple transfer learning when the external text data is available. Our final system drastically reduces the performance gap from the hybrid systems.