In 2020, the National Institute of Standards and Technology (NIST), in cooperation with the Intelligence Advanced Research Project Activity (IARPA), conducted an open challenge on automatic speech recognition (ASR) technology for low-resource languages on a challenging data type — conversational telephone speech. The OpenASR20 Challenge was offered for ten low-resource languages — Amharic, Cantonese, Guarani, Javanese, Kurmanji Kurdish, Mongolian, Pashto, Somali, Tamil, and Vietnamese. A total of nine teams from five countries fully participated, and 128 valid submissions were scored. This paper gives an overview of the challenge setup and procedures, as well as a summary of the results. The results show overall high word error rate (WER), with the best results on a severely constrained training data condition ranging from 0.4 to 0.65, depending on the language. ASR with such limited resources remains a challenging problem. Providing a computing platform may be a way to level the playing field and encourage wider participation in challenges like OpenASR.
Cite as: Peterson, K., Tong, A., Yu, Y. (2021) OpenASR20: An Open Challenge for Automatic Speech Recognition of Conversational Telephone Speech in Low-Resource Languages. Proc. Interspeech 2021, 4324-4328, doi: 10.21437/Interspeech.2021-1930
@inproceedings{peterson21_interspeech,
author={Kay Peterson and Audrey Tong and Yan Yu},
title={{OpenASR20: An Open Challenge for Automatic Speech Recognition of Conversational Telephone Speech in Low-Resource Languages}},
year=2021,
booktitle={Proc. Interspeech 2021},
pages={4324--4328},
doi={10.21437/Interspeech.2021-1930}
}