Publication Details
Transfer Learning Of Language-independent End-to-end ASR With Language Model Fusion
CHO, J.
Baskar Murali Karthick, Ing., Ph.D.
KAWAHARA, T.
Watanabe Shinji
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.
@inproceedings{BUT160002,
author="INAGUMA, H. and CHO, J. and BASKAR, M. and KAWAHARA, T. and WATANABE, S.",
title="Transfer Learning Of Language-independent End-to-end ASR With Language Model Fusion",
booktitle="Proceedings of ICASSP",
year="2019",
pages="6096--6100",
publisher="IEEE Signal Processing Society",
address="Brighton",
doi="10.1109/ICASSP.2019.8682918",
isbn="978-1-5386-4658-8",
url="https://ieeexplore.ieee.org/document/8682918"
}