Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Arabic-first generative speech recognition — Audar-ASR-V1 (Flash + Turbo). #1 on the Open Universal Arabic ASR Leaderboard. Model cards, benchmarks & inference.
| Date | Stars |
|---|---|
| 2026-07-24 | 558 |
| 2026-07-25 | 559 |
| 2026-07-28 | 559 |
| 2026-07-30 | 559 |
| 2026-07-31 | 561 |
| 2026-08-06 | 561 |
Today
— stars today
This week
+2 stars this week
This month
— stars this month
Momentum
4.0
growth rate 0.36%/day
<div align="center">
# Audar-ASR-V1
### Arabic-first generative speech recognition — state of the art on dialectal Arabic.
**From Arabic to the world.**
[](benchmarks/README.md)
[](https://huggingface.co/audarai)
[](report/Audar-ASR-V1-Technical-Report.pdf)
[](LICENSE)
[](https://www.audarai.com)
<p>
<a href="#-models"><b>Models</b></a> ·
<a href="#-quickstart"><b>Quickstart</b></a> ·
<a href="#-benchmarks"><b>Benchmarks</b></a> ·
<a href="#-downloads"><b>Downloads</b></a> ·
<a href="#-licenses"><b>Licenses</b></a> ·
<a href="https://www.audarai.com"><b>Audar API</b></a>
</p>
</div>
---
**Audar-ASR-V1** is a family of Arabic-first, generative speech-recognition models from
[AudarAI](https://www.audarai.com). 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](https://www.audarai.com/license/audarai-open-license-v1.0/) | [AudarAI Community v1.0](https://www.audarai.com/license/audarai-community-license-v1.0/) |
| **Download** | **[🤗 audarai/Audar-ASR-V1-Flash](https://huggingface.co/audarai/Audar-ASR-V1-Flash)** | **[🤗 audarai/Audar-ASR-V1-Turbo](https://huggingface.co/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)
```bash
pip install -r examples/requirements.txt
python examples/transcribe.py clip.wav # Arabic (auto-dialect)
python examples/transcribe.py english.wav --lang en
```
```python
# 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)
```bash
# Build a recent llama.cpp with Qwen3-ASR support, then:
./examples/gguf_infer.sh clip.wav turbo # or: flash
```Excerpt of 9,868 characters
Read on GitHub
6
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:72f45bb626184d39, topic:speech-recognition, topic:asr, topic:speech-to-text