SomaliBench

Somali speech benchmark

AI still gets Somali wrong. We are building the test that shows where.

SomaliBench collects consented Somali speech, transcribes it carefully, and uses it to score AI models against a private test set.

Sign-ups are not open yet.

How your voice is protected

What this is

SomaliBench is a project that records Somali speakers, with their permission, to measure how well AI understands the language.

  • Consented recordings

    Every speaker reads and signs a consent form before anything is recorded.

  • Careful transcription

    Each recording is written out carefully, so the answer key is one you can trust.

  • A private test set

    AI models are scored on audio that is never published, so the scores stay fair.

Why it matters

Somali is easy to get wrong, and hard to check.

Voice assistants, captions and translation tools are mostly built and tested on other languages. Somali speakers meet the mistakes every day: in names, in numbers, and in everyday words.

  • Dialects differ across Somalia, Somaliland, Djibouti, Ethiopia, Kenya and the diaspora.
  • Numbers, dates and money amounts need to be exactly right.
  • Many people mix Somali with English, Arabic or Swahili in the same sentence.

Who can take part

Somali-speaking adults, anywhere.

You must be 18 or older and speak Somali. Any dialect is welcome. You do not need special equipment: a phone is enough.

  • Age 18 or older
  • Speak Somali, in any dialect
  • Live anywhere in the world

How it works

Three steps.

  1. Sign up

    Create an account and tell us a little about your Somali: where you are from and where you learned it.

  2. Agree to the consent form

    Read the consent form in plain words. You choose what you agree to. Nothing is recorded before you sign.

  3. Talk

    Read short sentences or have a conversation. Recording is added in a later release.

How your voice is protected

Your voice, your choice.

Consent first
Nobody is recorded until they have read and signed the consent form.
No public audio
Recordings are never published. They are stored privately and shared only as the consent form describes.
No names on audio
Recordings carry a speaker ID. Your name, email and phone number are kept apart and are not shown to transcribers.
Withdraw any time
You can ask to withdraw from your settings page. New exports stop using your recordings straight away.
Read the privacy page

Help Somali be heard properly.

Joining takes a few minutes. You can read the full consent form before you decide.

Sign-ups are not open yet.