Central Kanuri Speech Dataset

You can access this dataset on HuggingFace.

Community-contributed Central Kanuri (kau) speech from Lake Chad Basin contributors, collected through the TWB Voice platform: 3,576 recordings across two workflows — 2,134 read-speech clips and 1,442 freeform spoken answers — with recording-level review status for every clip.

Languages

• knc — Central Kanuri

Dataset Structure

Recordings are embedded in the Parquet shards under data/<flow>/, one row per clip, in the audio column (HuggingFace Audiofeature). Select a workflow with config= and a review status with split=:

from datasets import load_dataset

ds = load_dataset("ciatech-frica/central-kanuri-speech-dataset", "Freeform", split="approved")

The two workflows were produced by different collection and quality-assurance pipelines and are kept apart deliberately. Read is read speech: the text read aloud is the prompt, so text is the prompt. Freeform is spontaneous speech answering a spoken question, so text is the answer and prompt is the question that was asked.

split is the platform’s own recording-approval status — approved, rejected or pending — not an ML train/dev/test partition. Nothing is withheld: every reviewed clip is published, and the status column plus the split name both tell you what survived review. rejected is the majority in the Read flow (1,020 of 2,134); for ASR training, filter on split="approved".

readback_candidates.tsv in the repository root indexes the clips discussed under Caveats below. Join it to this dataset on filename.

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)

Freeform

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

Collection ran on the TWB Voice platform operated by CLEAR Global, with CIATECH Africa leading implementation and community engagement across the Lake Chad Basin. Two workflows were run:

Read speech. Participants read predefined Central Kanuri prompts aloud.

Freeform speech. Participants heard a spoken Central Kanuri question and answered it spontaneously. Freeform clips then passed through recording review, transcription, transcription review, and correction where needed.

Contributors were onboarded and given recording guidance; recordings were reviewed by validators, and freeform recordings were transcribed and checked by transcribers. 24 contributors appear in the released metadata, across male and female speakers, recording predominantly in Nigeria.

Quality Control

Every clip carries the platform’s recording-approval status, and this release publishes all of them — nothing is withheld, so split and recording_approval_status are the filter to use.

595 Freeform recordings have no transcription record. 595 audio files were not transcribed and left as is.

2,982 rows carry a transcript: all 2,134 Read rows, where the transcript is the prompt, and 848 transcribed Freeform rows.

Ethical Considerations

• All contributors were community members who consented to their recordings being collected and published openly.
• Voice recordings are inherently biometric. Do not attempt to re-identify contributors, do not use this data for speaker identification or biometric profiling, and do not combine it with other data to identify individuals.
• Respect contributor privacy and comply with the licence conditions.

Caveats

• Some Freeform clips are read speech, not spontaneous speech. Contributors were asked a spoken question and a number read it back instead of answering, which would teach a spontaneous-ASR model to repeat its input. 192 of the 1,443 Freeform rows are candidates, listed in readback_candidates.tsv. No split or column marks them, so filter on that file. The list is a lower bound: a human reviewer who listened to the set reports most of it is read-back, so expect more read-backs to remain outside these rows.
• language_variant is a speaker demographic, like gender, not a label on the clip. It is the contributor’s own free-text answer at onboarding, and the recordings are Central Kanuri throughout — everyone was recorded on the same prompt pool. 6 of 24 speakers report a variety other than Central, so expect accented or non-native Central Kanuri from them. Kanembu and Suwurtiare Kanuri dialects, not separate languages.
• split is a review status, not a data partition. There is no train/dev/test split here. Make your own partition, speaker-disjointly: speaker_id groups clips by voice and there are only 24 speakers.
• 65 rows have no prompt_id. 555 were recovered by matching prompt text; 65 use prompts carrying no id anywhere in the corpus. prompt_id_source flags which rows are affected. This is a metadata gap, not an audio problem.
• 3,577 rows, 3,576 files: 6a3a88c4a2e82.wav is transcribed twice. Both rows are published. is_canonical_transcription marks the later one to prefer (true vs false) and is null on Read and untranscribed rows. Deduplicate on (flow, filename) if you need one row per clip.
• Audio was captured on contributor devices, not in a studio. Expect channel noise, reverberation, varying levels and sample rates.
• Scale and coverage. 3,576 clips from 24 contributors, one region and one linguistic setting. This is not representative of Kanuri as a whole, the two flows are not interchangeable for modelling, and the corpus is unlikely to be sufficient alone for training an end-to-end ASR system.

Related datasets

For ~50 hours of Kanuri read speech from the same platform, see TWB Voice 1.0.

License

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

For commercial licensing, contact contact@ciatech.ng.

Citation

@dataset{grema2026central_kanuri_speech,
  title     = {Central Kanuri Speech Dataset},
  author    = {CIATECH Africa and CLEAR Global},
  year      = {2026},
  publisher = {Hugging Face},
  license   = {CC BY-NC 4.0}
}

Contact

CIATECH Africa, project lead Umar Grema. CLEAR Global (TWB Voice platform) — tech@clearglobal.org.

Acknowledgments

CIATECH Africa and CLEAR Global acknowledge the Kanuri-speaking community members who contributed the recordings, and the language experts, validators, transcribers, coordinators and other contributors who supported the collection and its quality assurance.

We also acknowledge the TWB Voice team, who provided the data-collection platform and workflow infrastructure on which this corpus was produced.