We present a radio browsing system developed on a very small corpus of annotated speech by using semi-supervised training of multilingual DNN/HMM acoustic models. This system is intended to support relief and developmental programmes by the United Nations (UN) in parts of Africa where the spoken languages are extremely under resourced. We assume the availability of 12 minutes of annotated speech in the target language, and show how this can best be used to develop an acoustic model. First, a multilingual DNN/HMM is trained using Acholi as the target language and Luganda, Ugandan English and South African English as source languages. We show that the lowest word error rates are achieved by using this model to label further untranscribed target language data and then developing SGMM acoustic model from the extended dataset. The performance of an ASR system trained in this way is sufficient for keyword detection that yields useful and actionable near real-time information to developmental organisations.
Cite as: Saeb, A., Menon, R., Cameron, H., Kibira, W., Quinn, J., Niesler, T. (2017) Very Low Resource Radio Browsing for Agile Developmental and Humanitarian Monitoring. Proc. Interspeech 2017, 2118-2122, doi: 10.21437/Interspeech.2017-880
@inproceedings{saeb17_interspeech,
author={Armin Saeb and Raghav Menon and Hugh Cameron and William Kibira and John Quinn and Thomas Niesler},
title={{Very Low Resource Radio Browsing for Agile Developmental and Humanitarian Monitoring}},
year=2017,
booktitle={Proc. Interspeech 2017},
pages={2118--2122},
doi={10.21437/Interspeech.2017-880}
}