Introduction

Arabic speech recognition just got a serious open contender. On October 15, MBZUAI’s Institute of Foundation Models (IFM) and G42’s Inception42 launched the Jais 3 family, and the audio half of that launch is Jais-ASR: open weights under Apache 2.0, built on Jais-3-4B, covering Modern Standard Arabic plus dialects from 19 countries. You can run it today with Transformers or serve it with the bundled vLLM plugin; quantized MLX, GGUF, FP8, and AWQ builds sit in the Jais Speech collection.

The rest of the Jais 3 launch is a family of Arabic-native language models (0.9B to 375B). This piece stays on the speech side.

What Shipped

One model for Arabic as people actually speak it

Jais-ASR is aimed at real conversations, not just news-anchor MSA. The card lists dialect codes for Algeria, Bahrain, Egypt, the UAE, Iraq, Jordan, Kuwait, Lebanon, Libya, Mauritania, Morocco, Oman, Palestine, Qatar, Saudi Arabia, Sudan, Syria, Tunisia, and Yemen. It also handles the code-switching you hear in everyday speech: English, French, Spanish, and Italian spans can stay in their original script.

A few controls make it practical for makers:

  • code_switch=True keeps non-Arabic words as spoken instead of transliterating them.
  • diacritics=True adds vowel marks, which helps with pronunciation, learning apps, and feeding cleaner text into Arabic TTS.
  • Every transcript comes back with a language or dialect label (for example msa, EG, or en), so you get dialect ID without a second model.
  • When there is no speech (music, noise), it returns [IFM-NoTranscript] with refused=True instead of making text up. That is a lovely guardrail for podcast and video pipelines.

The numbers

On the community Open Universal Arabic ASR Leaderboard (six test sets covering read speech, broadcast, and conversational dialects), IFM reports a macro-average of 22.85% WER and 8.37% CER. That is ahead of Audar-ASR-V1-Turbo (23.17% WER), Speechmatics Melia 1 (23.51%), ElevenLabs Scribe v2 (24.54%), and open models like Qwen3-ASR-1.7B and Voxtral-Small-24B on the same sheet. It is not a clean sweep: on the Saudi broadcast set (SADA), Audar-ASR still wins. The two commercial APIs were measured by IFM itself, so treat those rows as vendor-reported until someone re-runs them.

What you need to run it

The bf16 weights are 11.4 GB, so plan on roughly 12 GB of GPU memory. Mac users can grab the 4-bit or 8-bit MLX builds, and the GGUF build is there for llama.cpp-style setups.

Try It

The quickest path is the Transformers pipeline:

pip install 'transformers==5.16.1' accelerate soundfile soxr

import torch
from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="IFM/Jais-ASR",
    trust_remote_code=True,
    device_map="auto",
    dtype=torch.bfloat16,
)
print(asr("speech.wav", max_new_tokens=256)["text"])

Want an OpenAI-style transcription endpoint for your app? The card ships a vLLM plugin (use a separate environment, since it pins different versions):

pip install 'vllm==0.23.0' 'transformers==5.5.4' accelerate soundfile soxr
hf download IFM/Jais-ASR --include "vllm_plugin/*" --local-dir Jais-ASR
pip install ./Jais-ASR/vllm_plugin
vllm serve IFM/Jais-ASR --trust-remote-code --reasoning-parser jais_asr_tag --max-model-len 2048

curl http://localhost:8000/v1/audio/transcriptions \
  -F file=@clip.wav \
  -F model=IFM/Jais-ASR

Why It Matters for Makers

Arabic voice work has been stuck between closed APIs and Whisper-style models that flatten dialects. Last week we covered TII’s Falcon ASR for Emirati Arabic; Jais-ASR goes wide instead, with one permissively licensed model across the whole region. For creators, that means subtitles for Egyptian comedy clips, Gulf podcasts, or Maghrebi interviews that keep the French and English bits intact, plus diacritized text you can hand straight to a voice model for dubbing. Test it on your own recordings before you trust the leaderboard, especially for long files, where you will want to raise max_new_tokens.

Conclusion

Jais-ASR is the most complete open Arabic speech recognizer we have seen so far: Apache 2.0 weights, 19 dialects, code-switching, diacritics, dialect labels, and a ready vLLM endpoint, all live on Hugging Face today. Grab a GPU with about 12 GB free, point it at a clip from your own corner of the Arabic-speaking world, and listen to how it does.

Original source:

Model card: IFM/Jais-ASR on Hugging Face · Collection: Jais Speech · Launch: Jais 3 blog

—Anabel ♡