🧭 What it is
Audar-ASR-V1-Turbo is Audar's proprietary Arabic speech-recognition model — the accuracy tier of
the Audar-ASR family. 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), and is
developed in-house through a proprietary Arabic training program:
- 🧱 Large-scale dialectal pretraining — 300,000+ hours of Arabic audio spanning MSA, Gulf,
Egyptian, Levantine and Maghrebi speech, code-switching, and diverse acoustic channels.
- 🎯 Dialect-targeted fine-tuning — hardness sampling and multi-task conditioning focused on proper
nouns, code-switching, and dialect-faithful orthography.
- 🧠 GRPO reinforcement-learning alignment — preference optimization against Arabic-native failure
modes (diacritization, code-switching, named-entity preservation, formatting) with trained native
annotators.
The result is
state-of-the-art dialectal Arabic ASR — the lowest average WER of any evaluated
system on the
Open Universal Arabic ASR Leaderboard. It transcribes MSA and every major Arabic
dialect, code-switched Arabic–English, and English, across
30 languages in total. For real-time,
edge, or high-throughput deployment, see the smaller
Audar-ASR-V1-Flash.
Distributed in the widely-supported Qwen3-ASR architecture format for turnkey tooling
(llama.cpp / GGUF). The model — data, training curriculum, and alignment — is Audar's.
Model summary
| Model | Audar-ASR-V1-Turbo — proprietary Arabic ASR (accuracy tier) |
| Task | Automatic speech recognition (audio → text) |
| Approach | Generative ASR — audio encoder + language-model decoder (audio-conditioned next-token prediction) |
| Training | 300k+ hrs dialectal pretraining → dialect-targeted SFT → GRPO alignment |
| Decoder parameters | 2,031,739,904 (2.03B) |
| Audio encoder parameters | 317,477,504 (0.32B) |
| Total parameters | 2,349,217,408 (2.35B, bf16) |
| Audio input | 16 kHz mono; 30 s context (longer audio is chunked/streamed) |
| Languages | Arabic (MSA + Gulf/Egyptian/Levantine/Maghrebi dialects) + English + 28 more |
| Runtime | GGUF / llama.cpp — CPU · GPU · edge |
| License | AudarAI Community License v1.0 |
📊 Benchmarks
Arabic dialectal ASR is hard — heavily dialectal, conversational, code-switched speech is the
frontier for every system. On the Open Universal Arabic ASR Leaderboard, Audar-ASR-V1-Turbo posts the
lowest average WER of any evaluated system on the full test sets — 24.7 %, best on four of the six —
and 3.55 % WER on CommonVoice-18 Arabic. The per-dataset development-protocol results (100
utterances/benchmark) are below.
Open Universal Arabic ASR Leaderboard — WER % (lower is better)
Per-dataset WER (%), development protocol (100 utterances/benchmark); baselines are the leaderboard's
published full-test scores. Best per column in bold. Authoritative full-test-set average: 24.7 %.
| System | CommonVoice-18 | MASC-clean | MASC-noisy | MGB-2 | SADA | Casablanca | Avg |
|---|
| Audar-ASR-V1-Turbo | 3.55 | 9.13 | 16.84 | 14.01 | 35.22 | 62.87 | 23.60 |
| ElevenLabs Scribe v1 | 5.74 | 9.87 | 19.78 | 15.15 | 40.87 | 66.93 | 26.39 |
| Qwen3-ASR-1.7B (base) | 10.86 | 15.07 | 21.12 | 29.21 | 50.54 | 85.25 | 35.34 |
| Whisper-Large-v3 | 17.83 | 24.66 | 34.63 | 16.26 | 55.96 | 71.81 | 36.86 |
Emirati Arabic
| Set | WER % | CER % |
|---|
| Emirati (Mixat, full 1,585-clip test) | 19.4 | 7.3 |
On Emirati, the real recognition error is ≈ 7.3 % — near-parity with spontaneous English — while the
residual up to 19.4 % WER is largely orthographic convention (near-miss spelling of the same
word, e.g. انتو↔انتوا, and Latin-vs-Arabic rendering of English loanwords), not misrecognition.
Measured on an internal dialectal validation sample
Same sample and harness as the Flash card
— useful for a direct Flash-vs-Turbo comparison (WER/CER %, N clips per set).
| Set (dialect) | N | WER % | CER % |
|---|
| SawtArabi (Gulf) | 23 | 13.7 | 2.7 |
| ArzEn (Egyptian ⇄ English code-switch) | 40 | 19.9 | 9.2 |
| MGB-3 (Egyptian broadcast) | 40 | 27.3 | 10.5 |
| Casablanca (Maghrebi / Moroccan Darija) | 40 | 61.9 | 28.6 |
Casablanca 61.9 WER ≈ the official leaderboard's 62.87 (reproduced in-house) — the numbers line up.
💻 GGUF inference (llama.cpp)
Turbo runs on llama.cpp via the multimodal (mtmd) path — a quantized decoder GGUF plus a
BF16 audio projector (mmproj). Build a recent llama.cpp (with Qwen3-ASR support), then:
1./llama-mtmd-cli \
2 -m Audar-ASR-V1-Turbo-Q8_0.gguf \
3 --mmproj mmproj-Audar-ASR-V1-Turbo.gguf \
4 --audio clip.wav \
5 -sys "فرّغ الكلام العربي التالي." \
6 --temp 0
⚠️ The audio projector (mmproj) must stay BF16 (its ClippableLinear is numerically
sensitive). The decoder quantizes normally.
Prefer a managed endpoint? The Audar-ASR family is also available via the
Audar API/SDK — streaming, speaker-attributed transcription, and
diarization, production-hosted.
GGUF variants
| File | Approx. size | Notes |
|---|
Audar-ASR-V1-Turbo-Q4_K_M.gguf | ~1.28 GB | Smallest; constrained hardware |
Audar-ASR-V1-Turbo-Q8_0.gguf | ~2.16 GB | Near-lossless (recommended) |
Audar-ASR-V1-Turbo.gguf (BF16) | ~4.07 GB | Full precision decoder |
mmproj-Audar-ASR-V1-Turbo.gguf | ~0.64 GB | BF16 audio encoder — required, keep BF16 |
🎙️ Real-time streaming
Audar-ASR streams via LocalAgreement-2: as audio arrives the trailing window is re-decoded each hop
and a word is committed only once two consecutive decodes agree on it — giving stable, low-latency
incremental output over the GGUF runtime. Audar's production realtime engine serves the same policy over
an OpenAI-Realtime-compatible WebSocket with model-based endpointing and ≥64 concurrent streams on a
single A100-80GB.
🌍 Languages, dialects & tasks
- 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.
- Task: transcription (audio → UTF-8 text), prompt-steerable for language and formatting.
Intended use & limitations
Intended use. Broadcast/media transcription, meeting & contact-center intelligence, voice agents,
captioning, and accessibility — cloud or on-prem.
Limitations.
- Maghrebi / Moroccan Darija (Casablanca) remains the hardest condition (~63 % WER) for all systems.
- Heavily code-switched telephony and low-SNR audio degrade accuracy relative to clean MSA.
- Long-form audio can drift on very long recordings.
- Not evaluated for, and must not be used for, covert speaker identification.
📜 License
Released under the
AudarAI Community License v1.0 — research and limited commercial use for
qualifying Community Entities; enterprise / large-scale / MaaS use requires an AudarAI Enterprise
License. See
audarai.com/license/audarai-community-license-v1.0.
Citation
1@misc{audar-asr-turbo-2026,
2 title = {Audar-ASR: Dialect-Aware Arabic Speech Recognition},
3 author = {AudarAI},
4 year = {2026},
5 note = {Audar-ASR-V1-Turbo},
6 url = {https://huggingface.co/audarai/Audar-ASR-V1-Turbo}
7}
About AudarAI
Leading Arabic-First Multilingual Audio Intelligence
AudarAI starts with Arabic — and expands to the world.
We are building advanced multilingual audio intelligence that helps individuals, enterprises, and
governments communicate across languages, cultures, and borders. By combining Arabic-first speech
technology with global multilingual AI, AudarAI transforms voice into understanding, interaction,
and connection.
Our work spans speech recognition, speech understanding, voice-enabled digital assistants,
human-computer interaction, and intelligent audio systems designed for real-world impact. From
empowering people to access technology in their native language to helping organizations
communicate globally, AudarAI is shaping a future where every voice can be heard, understood, and
connected.
Arabic-first. Multilingual by design. Human-centered at heart.
© 2026 AUDARAI PTE. LTD. · Licensed under the AudarAI Community License v1.0
-brightgreen){kind=icon}
