Transfer Learning Of Language-independent End-to-end ASR With Language Model Fusion

end-to-end ASR, multilingual speech recognition, low-resource language, transfer learning

This work explores better adaptation methods to low-resource lan-guages using an external language model (LM) under the frame-work of transfer learning. We first build a language-independentASR system in a unified sequence-to-sequence (S2S) architecturewith a shared vocabulary among all languages. During adaptation,we performLM fusion transfer, where an external LM is integratedinto the decoder network of the attention-based S2S model in thewhole adaptation stage, to effectively incorporate linguistic contextof the target language. We also investigate various seed models fortransfer learning. Experimental evaluations using the IARPA BA-BEL data set show that LM fusion transfer improves performanceson all target five languages compared with simple transfer learningwhen the external text data is available. Our final system drasticallyreduces the performance gap from the hybrid systems.