TICO-19 Speech – Congolese Swahili

Access this dataset on HuggingFace.

The dev and test portions of the TICO-19 benchmark read aloud in Congolese Swahili (swc): 2,992 clips (868 dev + 2,124 test) covering COVID-19 / medical-domain sentences, recorded in 2021 by CLEAR Global (formerly Translators without Borders) voice talent. Published as a single repository with two splits.

Languages

• swc — Swahili (Congo)

Dataset Structure

Recordings are embedded in the Parquet shards under data/read/, one row per clip, in the audio column (HuggingFace Audiofeature, mono 48 kHz MP3). The dev and test splits are two shard groups in the same repository; select one with split=.

from datasets import load_dataset

ds = load_dataset("CLEAR-Global/TICO-19-swc-speech", "read", split="test")

Each sentence is recorded by two voices, so every sentence_id appears in two rows within a split (once per session).

Statistics

read

Clips and hours by split

Hours by approval status

Demographic distributions below are computed on approved recordings only.

Hours by gender (approved)

Speakers by gender (approved)

Dataset Creation

Data Collection

The TICO-19 benchmark paragraphs (English, CMU origin) were translated through the pipeline English → French → Congo Swahili by the TICO-19 collaboration, then read aloud in Congo Swahili by CLEAR Global voice talent between February and April 2021. Audio is mono, 48 kHz MPEG Layer III.

This is the Congo Swahili rendering of the TICO-19 dev and test sets, where it served as the dev and evaluation data of the IWSLT 2021 Low-Resource Speech Translation shared task (Congolese Swahili→French). The corresponding training set (5,000 clips) is published separately as CLEAR-Global/Gamayun-swc-voice-minikit.

Quality Control

Recordings were quality-checked in-house by CLEAR Global for audio quality and for agreement with the text read. All clips in this release passed that review, so every row is published (approved); no pending or rejected rows are included, and no recordings were removed for offensive content or PII.

Speakers are identified only by anonymized numeric session IDs. Session-to-gender mapping comes from the delivery metadata: sessions 127, 128 and 130 are female; 125, 126 and 129 are male; 86 is female and 87 is male.

Ethical Considerations

• All speakers were professional voice talent who consented to recording and to open publication of the recordings.
• Speakers are identified only by anonymized numeric session IDs; no personal or otherwise identifying data is included.
• No attempts should be made to identify individual speakers.
• The data should be used responsibly, for research and educational purposes.
• The underlying TICO-19 benchmark text is published by its authors under CC0; see https://tico-19.github.io/.

Caveats

• The recordings were captured on mobile phones and are not studio quality; expect channel noise, reverberation and varying levels.
• Per the IWSLT 2021 low-resource task rules, the TICO-19 data was forbidden for training — it was used only for dev/validation and final evaluation. Please treat it as dev/eval data, not as additional training material, when comparing against IWSLT 2021 results.
• Only the Congo Swahili (target) side of TICO-19 is included here; the original French and English benchmark audio is not part of this release.
• The test split is not balanced: sessions 125, 126, 127 and 128 cover 2,050 clips, while sessions 129 and 130 contribute 37 clips each.
• Sentences are COVID-19 / medical-domain content; the corpus is small and topic-narrow.

License

CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International) (SPDX: cc-by-nc-4.0)

Citation

@inproceedings{anastasopoulos2020tico19,
  title     = {{TICO-19}: the Translation Initiative for COvid-19},
  author    = {Anastasopoulos, Antonios and Cattelan, Alessandro and Dou, Zi-Yi and Federico, Marcello and Federman, Christian and Genzel, Dmitriy and others},
  booktitle = {Proceedings of the 1st Workshop on NLP for COVID-19 (Part 2) at EMNLP 2020},
  year      = {2020},
  url       = {https://aclanthology.org/2020.nlpcovid19-2.5/},
  doi       = {10.18653/v1/2020.nlpcovid19-2.5}
}

@inproceedings{anastasopoulos2021findings,
  author    = {Anastasopoulos, Antonios and Bojar, Ond{\v{r}}ej and Declerck, Jind{\v{r}}ich and others},
  title     = {Findings of the {IWSLT} 2021 Evaluation Campaign},
  booktitle = {Proceedings of the 18th International Conference on Spoken Language Translation (IWSLT 2021)},
  year      = {2021},
  url       = {https://aclanthology.org/2021.iwslt-1.1/}
}

Contact

CLEAR Global — tech@clearglobal.org

Acknowledgments

Recorded by CLEAR Global voice talent in the Democratic Republic of the Congo. The TICO-19 benchmark was produced by the TICO-19 collaboration; see https://tico-19.github.io/ for the benchmark documentation and its CC0 text release. The corpus was assembled for the IWSLT 2021 Low-Resource Speech Translation shared task, organised with Antonios Anastasopoulos and the IWSLT team.