Skip to content
BlogAI

AI4 min read

Darja, French and English in one sentence: AI can finally transcribe it

A new open AI model transcribes Algerian Darja, even when it switches to French or English. What it can do, its limits, and three uses that make sense.

Amine KHDev, Tradix Technology
00:00-04:40

The problem every Algerian company knows

Listen to a sales call in an Algerian company. It starts in Darja, the product name is in French, a technical term comes in English, and the price is given in Arabic numbers, all in one sentence. Now imagine asking a transcription tool to write that down.
Until now, the result was usually unusable. Most speech-recognition tools are trained on English, on Modern Standard Arabic, or on Gulf and Egyptian dialects. Algerian Darja, with its French words and its own pronunciation, fell between the cracks. So the voice notes stayed voice notes: listened to once, then lost.

What was released

Jais Speech (published as Jais-ASR) was released by IFM, the team behind the Jais Arabic language models, in October 2026. According to its documentation:
  1. It understands Modern Standard Arabic and dialects from 19 countries. Algerian is explicitly on the list, along with Moroccan, Tunisian, Libyan and Mauritanian.
  2. It handles switching to French, English, Spanish or Italian in the middle of a sentence.
  3. It can detect which dialect is being spoken.
  4. It is released under the Apache 2.0 licence: free, including for commercial use.
  5. It has about 6 billion parameters and runs on a single graphics card with about 12 GB of memory, with smaller versions available.
Its makers report the lowest average error rate on the main public Arabic speech-recognition benchmark, ahead of the other open models compared and of two commercial services. That's a strong claim, but it comes from the vendor, and the same table shows other models doing better on some test sets.

Why "open" matters more than "best"

For me, the most interesting part isn't the benchmark. It's the licence.
Customer calls, negotiations, HR interviews and management meetings are among the most sensitive data a company has. Sending them to a foreign cloud service to be transcribed is a real problem, both for confidentiality and under Algeria's personal-data law. An open model changes the equation: the recordings can stay on a server in your building or in Algeria, and nobody else ever hears them.
The second point is cost. A cloud service charges per minute, forever. A model on your own server costs a machine and some setup, then nothing per minute, which matters if you process hundreds of calls a day.

Before you rely on it

I'd be careful about one thing: a benchmark is not your customers. Algerian Darja varies a lot between Algiers, Oran, Constantine and the South, and the test sets used to measure these models contain little of it. The only way to know whether it works for you is to try it on your own recordings.
Test it the way it will be used: real calls, from different regions, with phone-quality sound and mixed languages. Twenty recordings checked by someone who speaks the dialect will tell you more than any leaderboard.

Three uses that make sense today

1. Customer calls into the CRM. Each call is transcribed and attached to the customer's record, so the next person who calls them knows what was said.
2. Sales voice notes into orders. A sales rep sends a voice note from the field; the text is extracted, and the order is pre-filled in the ERP for someone to confirm.
3. Meeting minutes. The meeting is recorded, transcribed, and summarised, in the language people actually spoke.
In all three cases, a person still checks the result. The AI saves the typing, not the judgement.

The real lesson

For years, AI tools were built for other languages, and Algerian companies had to adapt to them. This is one of the first signs of the opposite: tools starting to adapt to the way people here actually speak. It's worth testing now, carefully, before your competitors do.
Share

Written by

Amine KH

Keep reading