End-to-end speech translation (E2E ST) has recently witnessed an increased interest given its system simplicity, lower inference latency and less compounding errors compared to cascaded one (speech recognition + machine translation). E2E ST model training, however, is often hampered by the lack of parallel data. Thus, we created CoVoST, a large & diverse multilingual speech-to-text translation corpus based on Common Voice (2019-06-12 release). It includes speeches in 11 languages (French, German, Dutch, Russian, Spanish, Italian, Turkish, Persian, Swedish, Mongolian and Chinese), their transcripts and English translations. We also provide an additional out-of-domain evaluation set from Tatoeba for 5 languages (French, German, Dutch, Russian and Spanish) into English.
Please check out our paper for more details and the VizSeq example for exploring CoVoST data.
- 2020-02-27: Example added for exploring CoVoST data with VizSeq
- 2020-02-13: Paper accepted to LREC 2020 (Oral)
- 2020-02-07: CoVoST released
-
Download the 2019-06-12 release of Common Voice (NOT the latest 2019-12-10 one from the web page) for speeches and transcripts:
-
Download translations for all the 11 languages, where
validated.<lang>_en.enare matched with the transcripts invalidated.tsv.
-
Download transcripts and translations and extract files to
data/tt/*. -
Download speech data:
python get_tt_speech.py --root <mp3 download root (default to data/tt/mp3)>| License | |
|---|---|
| CoVoST data | CC0 |
| Tatoeba sentences | CC BY 2.0 FR |
| Tatoeba speeches | Various CC licenses (please check out data/tt/tatoeba_s2t.<lang>_en.<lang>_lic) |
| Anything else | CC BY-NC 4.0 |
Please cite as
@misc{wang2020covost,
title={CoVoST: A Diverse Multilingual Speech-To-Text Translation Corpus},
author={Changhan Wang and Juan Pino and Anne Wu and Jiatao Gu},
year={2020},
eprint={2002.01320},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
Changhan Wang (changhan@fb.com), Juan Miguel Pino (juancarabina@fb.com), Jiatao Gu (jgu@fb.com)