Audar-ASR-V1

Arabic-first generative speech recognition — state of the art on dialectal Arabic.

From Arabic to the world.

Models · Quickstart · Benchmarks · Downloads · Licenses · Audar API

---

Audar-ASR-V1 is a family of Arabic-first, generative speech-recognition models from AudarAI. It recasts transcription as audio-conditioned next-token prediction over a unified text vocabulary — a language-model decoder rather than a CTC or transducer objective. It is built on a permissively-licensed open-weight audio-LLM foundation, then adapted in-house through 300,000+ hours of labeled audio (primarily Arabic, plus English) and a four-stage curriculum that ends in KTO preference alignment from native Arabic annotators — the contribution is the adaptation, not the foundation. It transcribes MSA and every major Arabic dialect (Gulf/Emirati, Egyptian, Levantine, Maghrebi), code-switched Arabic–English, and English — 30 languages in total.

On the Open Universal Arabic ASR Leaderboard, the Turbo tier posts the lowest average WER and CER of any evaluated system (rank #1 of 36), and the 0.78 B Flash tier is the strongest small model on the board (#11) — beating models several times its size. This repository is the developer hub: model pointers, benchmarks, and copy-paste inference for both tiers.

🧩 Models

	Audar-ASR-V1-Flash	Audar-ASR-V1-Turbo
Tier	Real-time · edge	Accuracy
Parameters	0.78 B (0.60 B decoder + 0.19 B encoder)	2.35 B (2.03 B decoder + 0.32 B encoder)
Runtimes	🤗 Transformers · GGUF (llama.cpp) · vLLM	GGUF (llama.cpp) · vLLM
Leaderboard	#11 of 36 · 33.31 % avg WER	#1 of 36 · 24.78 % avg WER
Best for	Live captioning, voice agents, on-device / offline	Lowest error on hard dialectal & long-form audio
License	AudarAI Open v1.0	AudarAI Community v1.0
Download	🤗 audarai/Audar-ASR-V1-Flash	🤗 audarai/Audar-ASR-V1-Turbo

Both share one architecture (Whisper-style 128-mel audio encoder → Qwen3 decoder, 30 s context) and one prompt interface, so you can develop against Flash and scale to Turbo without code changes.

⚡ Quickstart

Weights download automatically from the Hugging Face repos above.

Transformers — Flash (Python)

pip install -r examples/requirements.txt
python examples/transcribe.py clip.wav            # Arabic (auto-dialect)
python examples/transcribe.py english.wav --lang en

# Or call the reference helpers directly:
from audar_asr import load_model, transcribe
model, proc = load_model("audarai/Audar-ASR-V1-Flash")   # HF repo id or local path
print(transcribe(model, proc, "clip.wav"))               # <= 30 s clip

GGUF — Flash or Turbo (llama.cpp, CPU/GPU/edge)

# Build a recent llama.cpp with Qwen3-ASR support, then:
./examples/gguf_infer.sh clip.wav turbo     # or: flash

./build/bin/llama-mtmd-cli \
  -m       Audar-ASR-V1-Turbo-Q8_0.gguf \
  --mmproj mmproj-Audar-ASR-V1-Turbo.gguf \
  --audio  clip.wav \
  -sys     "فرّغ الكلام العربي التالي." \
  --temp 0

⚠️ The audio projector (mmproj) must stay BF16 — its ClippableLinear is numerically sensitive. The decoder GGUF quantizes normally (Q4_K_M / Q8_0 / BF16 all published on Hugging Face).

vLLM (GPU serving, OpenAI-compatible)

Both tiers serve on vLLM, which implements the Qwen3-ASR architecture natively — no custom serving code.

• Turbo serves a 4-bit W4A16 compressed-tensors build in the model repo's vllm-w4a16/ folder (~2.6 GB; near-BF16 accuracy, ≈ +1 pp CER).
• Flash serves directly from its bf16 model.safetensors — no quant step (~1.6 GB, lossless, FLEURS AR/EN CER ≈ 3.3 %). Flash prefixes raw output with a language <Lang><asr_text> tag — strip it client-side (re.sub(r"^\s*language\s+[A-Za-z]+\s*(?:<asr_text>)?\s*", "", text)).

./examples/vllm_serve.sh turbo    # or: flash — builds an audio-enabled vLLM image + serves on :8000

The stock vLLM image omits audio codecs, so the helper adds them (av librosa soundfile). Once the server is up, transcribe over the OpenAI API — either POST /v1/audio/transcriptions (multipart) or /v1/chat/completions with an input_audio part and the Arabic system prompt:

curl -s http://localhost:8000/v1/audio/transcriptions \
  -F model=audar-asr-v1-turbo -F file=@clip.wav -F temperature=0

Long-form & realtime streaming

python examples/transcribe_long.py meeting.wav   # arbitrary length (30 s chunking)
python examples/stream.py long.wav               # LocalAgreement-2 incremental output

Streaming commits a word only once two consecutive sliding-window decodes agree on it, so committed text is stable and never rewrites. This is the local reference policy; Audar's production engine serves the same policy with sub-250 ms latency over an OpenAI-Realtime-compatible WebSocket — see the Audar API.

📊 Benchmarks

Full-test-set results on the Open Universal Arabic ASR Leaderboard, same harness and normalizer as the public board (calibrated within 0.03 pp). Top of the table:

#	Model	Avg WER	Avg CER
1	Audar-ASR-V1-Turbo	24.78	9.49
2	CohereLabs/cohere-transcribe-arabic-07-2026	25.87	11.80
3	omnilingual-asr/omniASR_LLM_7B	28.32	12.52
…
11	Audar-ASR-V1-Flash (0.78 B)	33.31	13.66
12	Qwen/Qwen3-ASR-1.7B	33.36	12.33

Full 36-system standings, per-dataset WER/CER, methodology, and machine-readable leaderboard.csv → benchmarks/.

Highlights: Turbo is #1 overall and best on dialect-heavy SADA and MGB-2; Flash (0.78 B) beats Qwen3-ASR-1.7B, Voxtral-Small-24B, and Whisper-large-v3. On in-house Gulf-Emirati long-form, Turbo hits 19.4 % WER / 7.3 % CER.

📥 Downloads

All weights and GGUF variants live on Hugging Face:

• Flash (Transformers + GGUF): https://huggingface.co/audarai/Audar-ASR-V1-Flash
• Turbo (GGUF): https://huggingface.co/audarai/Audar-ASR-V1-Turbo
• Full model collection: https://huggingface.co/audarai

Prefer a managed, production-hosted endpoint (streaming, speaker-attributed transcription, diarization)? See the Audar API.

🌍 Languages & capabilities

• Primary: Arabic — MSA and dialectal (Gulf/Emirati, Egyptian, Levantine, Maghrebi), plus code-switched Arabic–English; emits dialect-faithful orthography from audio alone.
• Also: English + 28 additional languages (30 total).
• Tasks: transcription (audio → UTF-8 text), prompt-steerable for language and formatting; long-form and realtime streaming via LocalAgreement-2.
• Not for covert speaker identification.

📜 Licenses

• This repository (example code, reference helpers, docs) — Apache-2.0 (LICENSE).
• Model weights — released under AudarAI model licenses (not Apache-2.0):
  • Audar-ASR-V1-Flash → AudarAI Open License v1.0 (commercial use, redistribution, and fine-tuning permitted).
  • Audar-ASR-V1-Turbo → AudarAI Community License v1.0 (research and limited commercial use for qualifying Community Entities).
• Enterprise / large-scale / model-as-a-service use may require an AudarAI Enterprise License — contact contact@audarai.com or visit audarai.com.

📄 Citation

Technical report — Audar-ASR-V1: A Multilingual, Arabic-First Generative Speech Recognition Foundation Model, Audar AI Team, 2026 — read the PDF.

@techreport{audar-asr-v1-2026,
  title       = {Audar-ASR-V1: A Multilingual, Arabic-First Generative Speech Recognition Foundation Model},
  author      = {Audar AI Team},
  institution = {AudarAI},
  year        = {2026},
  url         = {https://github.com/AudarAI/Audar-ASR-V1}
}

---

Leading Arabic-First Multilingual Audio Intelligence

AudarAI starts with Arabic — and expands to the world.

Arabic-first. Multilingual by design. Human-centered at heart.

🌐 www.audarai.com · 🤗 Hugging Face · GitHub · contact@audarai.com

© 2026 AUDARAI PTE. LTD.