Zarma
Zarma from the ground up
What does it take to build speech technology for a language that almost none exists for?
The first open, consented speech resources and honest baselines for Zarma, spoken by about six million people along the Niger. Zarma is spoken by about six million people along the Niger River, and it is one of the main languages of Niamey.
Orthography
A normaliser vetted by Zarma writers outside the founding team.
A test set first
A few consented hours, verified twice and frozen.
Baselines
Existing systems on that test set, scored strict and folded.
Zarma is spoken by about six million people along the Niger River, and it is one of the main languages of Niamey. It has four tones, long and short vowels, nasal vowels and letters like ɲ and ŋ. As far as we can find, it has no Common Voice collection, no FLEURS test set and no adapter in Meta’s widely used MMS recogniser. The largest speech resource, a Niger radio corpus described in 2026, has about sixty hours of verified Zarma transcription.
This programme builds Zarma speech technology from the ground up, and in the right order. First a written convention that Zarma speakers recognise, aligned with Niger’s official orthography and honest about how people actually spell. Then a small, consented, carefully transcribed test set, so that every later claim has something to stand on. Then recordings and models: recognition first, then synthesis, then conversation.
Zarma is also our test of method. We built our Hausa work so that a new language is added through a separate language pack, not by rewriting the system. Zarma will show whether that holds, how long it takes, and what does not transfer.
We want to do this with Zarma speakers, writers and the researchers already building Zarma resources, not around them.
What we do not claim yet
- Namu has no Zarma data, model or measurement yet. This programme is starting.
- Figures about existing Zarma corpora are their authors’ figures, not ours.
- Zarma has an official orthography (1999); everyday spelling varies. We do not say it has no writing system.
QuestionsWhat this programme sets out to answer
- 01
What written convention should Zarma transcripts follow, and how should recognition be scored when more than one spelling is accepted?
- 02
How well do multilingual systems that list Zarma actually recognise it on a clean, verified test set?
- 03
How many hours of verified Zarma does a usable recogniser need, and does Hausa data help or hurt?
- 04
Can a language pack add Zarma without rewriting the system, and what does that cost?
ProgressDates appear only on finished steps
- NextOrthography and normaliserVetted by at least two Zarma writers outside the founding team.
- NextA test set of a few hoursConsented, verified twice, frozen.
- NextBaselinesExisting systems on that test set, scored strict and folded.
- NextTraining dataRadio-corpus access if granted, plus consented collection including Niamey code-mixed speech.
- NextZarma recognition and synthesis v0With native listening evaluation.
What we will release
- A Zarma text-normalisation spec with test cases
- Zarma test v0, with a data statement
- A public baseline report
- Paper: adding a language by language pack
Work with us on Zarma
We are looking for Zarma writers and linguists, researchers building Zarma resources, and speakers in Niamey and along the river.
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.
Tankaɗe
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.
Maʼauni
Test sets, text normalisation, listening protocols and benchmarks that make a Hausa or Zarma result mean something.