Tankaɗe
From raw audio to trustworthy data
Which choices in cleaning speech data change what a model learns?
The methods that turn radio archives, field recordings and crowd audio into Hausa and Zarma data a model can learn from and a reader can trust. Most Hausa audio in the world was never made for machines.
Segment
Cut long broadcasts and field recordings into utterances.
Agree
Keep what two recognisers agree on, align it word by word.
Repair
Restore hooked letters and spelling, and tag every licence.
Most Hausa audio in the world was never made for machines. It is radio programmes with music under the news, phone-in interviews, sermons, people reading sentences into cheap phones. Before any of it can teach a model, someone has to decide what is speech, who is speaking, what was said, how it should be spelled, and whether it may be used at all.
Tankaɗe is our research into that work. We treat it as research, not cleaning up, because every choice changes what a model learns. Cut segments at the wrong place and turn-taking disappears. Accept a transcript that drops the hooked letters ɓ, ɗ and ƙ, and the model learns that they do not matter. Enhance noisy audio with a generative model, and you train on sound that no one spoke.
Our pipeline segments and diarises broadcast audio, filters transcripts by the agreement of two independent recognisers, aligns every word in time, repairs and standardises Hausa spelling, removes music and rebroadcasts, and tags every derived file with its source’s licence. We measure what each step keeps and what it throws away, and we check the thresholds against human transcripts.
We will release the tools, the thresholds and the yields, so that others working on African languages can sift their own audio, and see where ours went wrong.
What we do not claim yet
- Our yields and thresholds are internal until they are published in a write-up with their conditions.
- The pipeline is not finished: word timings are provisional until hand labels exist.
- We do not redistribute third-party broadcast content or any source beyond its licence.
The life of an utterance Recorded, segmented, aligned, normalised and checked against the speaker’s consent.
QuestionsWhat this programme sets out to answer
- 01
How should agreement-based pseudo-labelling be calibrated when the written references themselves are inconsistent?
- 02
Can real conversational structure be recovered from mono broadcast audio, and how wrong is it compared with true two-channel recordings?
- 03
What does generative enhancement do to tone, hooked consonants and speaker identity?
- 04
How should licence and consent restrictions travel through tokens, alignments and checkpoints so that a forbidden use is impossible?
ProgressDates appear only on finished steps
- Done · Oct 2026Broadcast and spontaneous-speech pipelines builtSegmentation, two-recogniser agreement, alignment, spelling repair and licence tags. Internal.
- In progressWrite up what existsThe pipelines and their yields, to the release standard.
- NextHuman checkA hand-labelled audit set of transcripts and word timings; thresholds re-validated.
- NextRelease the toolWith the Hausa language pack and test cases.
- NextRelease derivativesPublic-domain broadcast transcripts and alignments, with a data statement.
- NextZarma packApply the tool to Zarma and report what broke.
What we will release
- An open curation pipeline with Hausa and Zarma packs
- Curated derivatives where licences allow
- A licence-propagation schema
- Paper: sifting Hausa broadcast audio — yields, thresholds and what they cost
Work with us on Tankaɗe
We are looking for Hausa transcribers to build the audit set, and groups curating audio in other African languages to compare methods with.
Get in touchMore research
All researchHira
Speech models that listen while they speak, built for Hausa and judged against how Hausa speakers actually take turns.
Rumbu
Hausa speech collected with consent and fair pay, from the people and places public datasets leave out.
Zarma
The first open, consented speech resources and honest baselines for Zarma, spoken by about six million people along the Niger.
Maʼauni
Test sets, text normalisation, listening protocols and benchmarks that make a Hausa or Zarma result mean something.