GLOBAL ALLIANCE FOR LIVING VOICES

Machines are learning to listen. Whose voices will they hear?

Today’s AI learned the world mostly in written English. Every other language reaches it through translation, and what written English cannot say disappears. The Awal Alliance brings countries and communities together to build AI that listens first, and to keep every voice where it belongs.

■ Tachelhit   ■ Darija

One voice, two languages, one sentence. Most machines would force it into a single written language. We keep it whole.

Awal (ⴰⵡⴰⵍ) means “the spoken word” in Amazigh: the living voice, and the word one gives as a pledge. A word from a language long held to be minor, to name an alliance that gives the word back to every language that was silenced.

WHY IT MATTERS

The next few years will decide, for a long time, which voices machines understand.

01 · What is being lost

Elders who carry tales, songs, remedies and the names of places are not recorded, and will not be recorded twice. When the last speakers fall silent, a whole way of cutting up the world goes with them.

02 · Why translation is a trap

When every word must pass through written English to be processed, what English cannot say is dropped. The maternal uncle becomes just “an uncle”. An oath becomes mere insistence. Generation after generation, heritage is poured into a single mould.

03 · What Inverse AI does instead

It starts from the voice, not the text. The machine adapts to the speaker, not the reverse. Speakers, not a text score, judge whether it works. And everything it produces can be traced back to the original voice.

Read the four stories →

THE CORE PRINCIPLE

Share the method and the tools. Protect the voices.

Like the W3C, the Alliance brings its members around one table to build shared standards and infrastructure. Unlike any data platform, it never asks anyone to hand over their voices.

What we share

  • The method: Inverse AI, from collection to evaluation
  • The tools: offline collection on a simple phone, annotation for languages without fixed spelling
  • The protocols: oral consent, in the speaker’s own language
  • The measures: speakers judge, results published by variety
  • The training: for collectors, archivists and researchers

What we protect

  • The voices: they belong to those who spoke and to their communities
  • The languages: processed in their own terms, never through a pivot language
  • The data: kept in its country; only models and results travel
  • The knowledge: each community sets its access, open, restricted or closed
  • The right to say no: consent can be withdrawn at any time

The gap is measurable.

146%

Word error rate of a leading speech recognition model on Tachelhit, and 76% on Moroccan Darija, without adaptation. It makes more errors than there are words.

Whisper large-v3 · Atouf et al., 2025

~1,000

Speakers of Haketia, the Judeo-Moroccan Spanish, left today, against some 30,000 around 1900. Most are elderly and scattered across four continents.

Jewish Language Project

~7,000

Languages spoken in the world today. Most are oral first, and most reach AI only through translation, if at all.

Ethnologue

BEING FOUNDED NOW

Three signatures bring the Charter to life.

The Charter enters into force when three founding members have signed it, including at least one government and one community of guardians. The Alliance starts in Morocco, where orality, multilingualism and urgency meet.

  • Emergency collection for Haketia, with its last speakers in Morocco, Montreal, Spain and Israel.
  • A LUMINA pilot on Darija and one Amazigh variety, with volunteer communities.
  • A founding core: a Moroccan university, a university abroad, a guardians’ association, a heritage institution and a government member.