Speech recognition that finally puts Arabic first

What it does Two tiers—Flash (0.78 B) for edge and Turbo (2.35 B) for accuracy—transcribe Modern Standard Arabic, every major dialect, code-switched Arabic–English, English, and 28 other languages. Both use the same architecture: a Whisper-style audio encoder feeding a Qwen3 decoder that treats transcription as next-token prediction. The repo holds model pointers, benchmark data, and reference inference scripts.

The interesting bit The authors are upfront that the base architecture is an off-the-shelf audio-LLM; their contribution is the adaptation. They threw 300,000+ hours of labeled audio at it, ran a four-stage training curriculum, and finished with KTO preference alignment from native Arabic annotators. The result is Turbo sitting at rank #1 on the Open Universal Arabic ASR Leaderboard and a sub-billion-parameter Flash model that beats models several times its size.

Key highlights

• Turbo posts a 24.78 % average WER on the Open Universal Arabic ASR Leaderboard (#1 of 36); Flash hits 33.31 % WER at #11, beating models several times its size.
• Flash runs fully offline via llama.cpp GGUF or Hugging Face Transformers; Turbo serves through vLLM with an OpenAI-compatible endpoint.
• Both tiers share one prompt interface and 30-second context window, so you can prototype on Flash and hot-swap to Turbo without touching code.
• Weights are released under custom AudarAI licenses (Open for Flash, Community for Turbo), while the repo code itself is Apache-2.0.
• Realtime streaming uses a LocalAgreement-2 policy that commits a word only after two consecutive sliding-window decodes agree, preventing mid-word rewrites.

Caveats

• The audio projector (mmproj) must stay in BF16 when quantizing GGUF builds; the README warns it is numerically sensitive.
• Turbo’s weights are released under the AudarAI Community License, which limits large-scale commercial use; enterprise redistribution requires a separate agreement.
• The stock vLLM image lacks audio codecs, so the provided helper scripts patch in av, librosa, and soundfile.

Verdict If you are building voice apps for Arabic dialects or edge devices in the Middle East, this is worth a hard look. If you need a fully open-weight model under a standard permissive license for unrestricted commercial redistribution, the custom licensing on the weights will give you pause.