This paper studies topic and keyword identification for languages in which we have no transcribed speech data. We adopt a transfer learning framework to transfer what is learned from rich-resourced languages (RRL) to low-resourced languages (LRL). Specifically, we propose that a convolutional neural network (CNN) trained as a topic classifier in an RRL learns features (hidden layer activations) that can be used for the same purpose in an LRL. The CNN observes acoustic features, RRL phones, or segment clusters generated by an unsupervised phone clustering system; its hidden layers are retained and its output layer re-trained from scratch on the LRL. Our results are compared with the state-of-the-art topic classification methods on cross-language ASR transcripts. We also discuss the successful detection of topic dependent keywords and the use of unsupervised learning based clusters in our approach for low-resourced language topic detection.
Cite as: Chen, W., Hasegawa-Johnson, M., Chen, N.F. (2018) Topic and Keyword Identification for Low-resourced Speech Using Cross-Language Transfer Learning. Proc. Interspeech 2018, 2047-2051, doi: 10.21437/Interspeech.2018-1283
@inproceedings{chen18e_interspeech,
author={Wenda Chen and Mark Hasegawa-Johnson and Nancy F. Chen},
title={{Topic and Keyword Identification for Low-resourced Speech Using Cross-Language Transfer Learning}},
year=2018,
booktitle={Proc. Interspeech 2018},
pages={2047--2051},
doi={10.21437/Interspeech.2018-1283}
}