Namu-Transcribe

Turn noisy Hausa audio into accurate, domain-aware transcripts.

Features

Namu-Transcribe is trained on Hausa the way people speak it: across dialects, over bad lines, and in rooms where other people are talking.

Accurate across eight Hausa dialects

Measured separately on each, because an average across Niger hides the varieties it does worst on.

MaradiZinderTahouaTessaouaKonniDogondoutchiMadaouaMirriah

It learns your words

Every trade has names of its own. Tell Namu the ones you use and it will stop guessing at them.

Presenter00:04

Barka da yamma. Wannan shi ne labarin karfe shida na yammacin yau.

Presenter00:11

Ruwan sama ya sauka a yankin Maradi da Tessaoua a daren jiya, kuma manoma na sa ran fara shuka a mako mai zuwa.

Correspondent00:23

A nan Zinder, kasuwar dabbobi ta bude da karfe bakwai. Farashin raguna ya ragu kadan idan aka kwatanta da makon jiya.

Made for real recordings

Built for imperfect conditions: a phone on a market stall, a radio feed, a room with three people talking at once.

A field seen through trees from a moving vehicle

Ina son sanin farashin gero a kasuwar yau.

Using the model

Hausa in, text out.

Ruwan sama ya sauka a yankin Maradi da Tessaoua a daren jiya, kuma manoma na sa ran fara shuka a mako mai zuwa.

00:00 00:00

Performance

Measured on Hausa, not on average

Benchmarks that quote one number across forty languages tell you very little about the one you speak. These are Hausa alone.

8.4%

Word error rate

Read speech, clean audio

14.1%

In the field

Market and street recordings

8

Dialects measured

Across Niger

11×

Faster than real time

Batch, single GPU

Domain termsSupported
TimestampsWord level
StreamingYes
Price$0.22 per hour of audio

Try Namu-Transcribe

Namu Playground

Record a line of Hausa, or drop in a file, and read it back.

Try in playground

Namu API

Transcribe at volume, with your own domain terms loaded.

Contact sales

All Namu models

Interpretation, speech synthesis and the end-to-end voice agent.

See all models