Training and validation data for the DSN 2026 AI Bootcamp competition Find the Language I Am Speaking. The task: given a short audio clip, predict which of 12 Nigerian languages is spoken, or other. The hidden test set is not part of this dataset.

data/train-*.parquet (24774 clips) and data/validation-*.parquet (4352 clips).
Columns: id, audio (16 kHz mono FLAC clips of 2 to 30 s, stored inside the files), label, label_id, speaker_id, source, duration.
Rows are sorted by label, so a query for one language reads only a small part of the data.
ATTRIBUTION.md, LICENSE.md: sources and their licenses. PROHIBITED_SOURCES.txt: corpora the hidden test set was drawn from.

Load it

from datasets import load_dataset
ds = load_dataset("Data-Science-Nigeria/dsn-2026-lid-train")                  # splits: train, validation
ds = load_dataset("Data-Science-Nigeria/dsn-2026-lid-train", streaming=True)  # without downloading everything first
print(ds["train"][0])   # {"audio": ..., "label": "yor", "speaker_id": ..., ...}

Only one language, with SQL (DuckDB reads just the matching parts of the shards):

SELECT label, count(*) AS clips, round(sum(duration) / 3600, 2) AS hours
FROM 'hf://datasets/Data-Science-Nigeria/dsn-2026-lid-train/data/train-*.parquet' GROUP BY label ORDER BY hours DESC;

SELECT id, speaker_id, duration FROM 'hf://datasets/Data-Science-Nigeria/dsn-2026-lid-train/data/train-*.parquet' WHERE label = 'ego';

Labels and size

Class	Name	Train h	Validation h	Train clips	Validation clips
`yor`	Yoruba	9.05	0.88	3160	330
`ibo`	Igbo	7.96	2.08	2592	704
`hau`	Hausa	8.31	1.42	2755	495
`ego`	Eggon	7.32	1.05	1631	222
`anw`	Annang	8.12	1.83	1912	387
`efi`	Efik	1.42	0.15	483	64
`ibb`	Ibibio	1.76	0.28	607	107
`idu`	Idoma	7.16	2.12	1649	450
`igl`	Igala	0.48	0.12	170	49
`nup`	Nupe	0.50	0.00	176	0
`tiv`	Tiv	4.27	0.63	1474	213
`urh`	Urhobo	9.22	1.57	3122	550
`other`	Other (non-target speech)	12.98	1.85	5043	781

Several languages have only a few hours of audio, so the data is strongly imbalanced. Use class weighting, oversampling and augmentation. Speakers are kept within one split where speaker information exists, so validation scores reflect unseen speakers.

What was changed from the original sources

Audio was resampled to 16 kHz mono and stored as FLAC. Recordings longer than 30 seconds were cut into pieces at quiet points. Labels were mapped to ISO 639-3 codes (other covers English, Nigerian Pidgin, other Nigerian languages and further languages). Exact duplicates were removed, and splits were rebuilt by speaker. The clips were then packed into Parquet files.

Sources and licenses

This dataset combines several sources, each under its own license. The license field is other on purpose: see LICENSE.md for the license of each source, and keep their attribution if you share anything derived from this data.

Per source

This dataset combines the sources below. Keep their attribution and license terms.

afrispeech_ipa: CC-BY-4.0 (stated; upstream audio terms not verified)
hypa_voices: Apache-2.0
omnilingual_asr: CC-BY-4.0

If you use the Omnilingual ASR Corpus portion in research, please cite:

@misc{omnilingualasr2025,
    title={{Omnilingual ASR}: Open-Source Multilingual Speech Recognition for 1600+ Languages},
    author={{Omnilingual ASR Team} and others},
    year={2025}, eprint={2511.09690}, archivePrefix={arXiv}, primaryClass={cs.CL},
    url={https://arxiv.org/abs/2511.09690}
}

Limitations

Some languages have few hours and few speakers; read recordings dominate some sources.
Audio from some sources may be read religious text, which can differ in style from everyday speech.
The dataset is meant for language identification; transcripts are not included.