namu
namu-ai.orgHear Aïr
Active

Hira

Full-duplex conversation in Hausa

Read the questionsSee progressCollaborate

Can a machine listen while it speaks, in Hausa, and take turns the way Hausa speakers do?

Speech models that listen while they speak, built for Hausa and judged against how Hausa speakers actually take turns. People do not talk in turns that wait politely for silence.

An ear

Turns Hausa speech into something a language model can read.

A brain

A multilingual model, whose weights we do not retrain, decides what to say.

A mouth

Built on Kyutai’s Moshi architecture, it speaks while it keeps listening.

People do not talk in turns that wait politely for silence. They answer before a sentence ends, say “to” and “eh” while the other person is still speaking, stop when interrupted, and stretch a greeting across many exchanges because the greeting is the point. Most voice systems ignore all of this: they wait, transcribe, think, then speak.

Hira is our research into full-duplex speech for Hausa: models that hear and speak at the same time. Our working system, Aïr, separates the problem in three. An ear turns Hausa speech into something a language model can read. A brain, a multilingual model whose weights we do not retrain, decides what to say. A mouth, built on Kyutai’s Moshi architecture, speaks while it keeps listening, so it can wait, stop or acknowledge.

We started there because our first attempt, teaching one model everything at once, learned the rhythm of conversation but not its sense. The questions now are harder ones. How do Hausa speakers time their turns? Which backchannels do they use, and where? Can a model learn this from tens of hours of real conversation when published systems used tens of thousands? And how do we measure it without borrowing English answers?

We will publish what we measure, including what does not work.

What we do not claim yet

  • Aïr’s live demo is not end-to-end speech-to-speech today: its hearing still goes through a transcript.
  • Aïr has not learned Hausa turn-taking. Its timing is a policy plus distributions from other languages until real Hausa conversation is measured.
  • No Aïr result is published yet. Internal numbers will appear with their conditions, after review.
  • The brain is a third-party multilingual model (McGill AfriqueQwen). We do not retrain its weights, and we did not train it.

Two voices, one timeline Full duplex means overlap, backchannels and interruptions. This is how the Lab will show them for every recorded conversation.

Lines from namu-ai.org’s Aïr page. Timing is illustrative until a consented two-speaker recording is published.

QuestionsWhat this programme sets out to answer

  1. 01

    What are the timing norms of Hausa conversation — gaps, overlaps, backchannel rate, placement and inventory — in Niger and in Nigeria, face to face and on the phone?

  2. 02

    Can a duplex model keep the turn-taking skills of an English base while it learns Hausa content from a separate brain, and how much real Hausa conversation does it need to sound Hausa rather than translated?

  3. 03

    Is the thinker–talker split (an ear, a brain and a mouth) a better trade than single-stream duplex models for low-resource languages, measured on relevance, latency and naturalness together?

  4. 04

    How should full-duplex behaviour be evaluated in a language with no turn-taking benchmark, and who judges it?

ProgressDates appear only on finished steps

  1. Done · Oct 2026Thinker–talker prototype runningEar, brain and mouth run as one live research chain. Internal; write-up to follow.
  2. In progressAïr speech-to-speech end to endTraining the ear on real phone audio so hearing no longer goes through a transcript.
  3. NextMeasure people firstRecord consented two-channel Hausa conversations in Niger and Nigeria; publish their timing statistics.
  4. NextA Hausa voice-activity-projection modelTest whether English or multilingual turn-taking models transfer, against a held-out Hausa set.
  5. NextHausa full-duplex benchmark v0Tasks scored against human Hausa conversation, with a judging protocol calibrated on native raters.
  6. NextNative conversational judgementAt least five raters across dialects, agreement reported.
  7. NextPublic demo and paperWith a system card and a voice-consent record.

What we will release

  • Hausa conversation timing statistics
  • A Hausa full-duplex benchmark
  • Paper: thinker–talker for low-resource full duplex, with the negative result that led to it
  • A disclosed research-preview demo of Aïr

Work with us on Hira

We are looking for paid Hausa speakers for two-person recordings, native-speaker raters across dialects, and researchers working on turn-taking.

Get in touch

Related reading

More research

All research

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.

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.