[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-93224":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":13,"contributorsCount":13,"subscribersCount":13,"size":13,"stars1d":13,"stars7d":13,"stars30d":15,"stars90d":13,"forks30d":13,"starsTrendScore":13,"compositeScore":16,"rankGlobal":10,"rankLanguage":10,"license":17,"archived":18,"fork":18,"defaultBranch":19,"hasWiki":20,"hasPages":18,"topics":21,"createdAt":10,"pushedAt":10,"updatedAt":32,"readmeContent":33,"aiSummary":34,"trendingCount":13,"starSnapshotCount":13,"syncStatus":35,"lastSyncTime":36,"discoverSource":37},93224,"Audar-ASR-V1","AudarAI\u002FAudar-ASR-V1","AudarAI","Arabic-first generative speech recognition — Audar-ASR-V1 (Flash + Turbo). #1 on the Open Universal Arabic ASR Leaderboard. Model cards, benchmarks & inference.","https:\u002F\u002Fwww.audarai.com",null,"Python",110,0,103,7,40.7,"Apache License 2.0",false,"main",true,[22,23,24,25,26,27,28,29,30,31],"arabic","arabic-nlp","asr","audar","dialectal-arabic","gguf","huggingface","llama-cpp","speech-recognition","speech-to-text","2026-07-22 04:02:08","\u003Cdiv align=\"center\">\n\n# Audar-ASR-V1\n\n### Arabic-first generative speech recognition — state of the art on dialectal Arabic.\n\n**From Arabic to the world.**\n\n[![Open-AR-ASR](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpen--AR--ASR-%231%20of%2036%20%C2%B7%2024.8%25%20avg%20WER-2ea44f)](benchmarks\u002FREADME.md)\n[![Models](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F🤗%20Models-Flash%20%2B%20Turbo-ffcc4d)](https:\u002F\u002Fhuggingface.co\u002Faudarai)\n[![Paper](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F📄%20Technical%20Report-PDF-blue)](report\u002FAudar-ASR-V1-Technical-Report.pdf)\n[![Code License](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fcode-Apache%202.0-6f42c1)](LICENSE)\n[![Website](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F🌐%20audarai.com-informational)](https:\u002F\u002Fwww.audarai.com)\n\n\u003Cp>\n\u003Ca href=\"#-models\">\u003Cb>Models\u003C\u002Fb>\u003C\u002Fa> ·\n\u003Ca href=\"#-quickstart\">\u003Cb>Quickstart\u003C\u002Fb>\u003C\u002Fa> ·\n\u003Ca href=\"#-benchmarks\">\u003Cb>Benchmarks\u003C\u002Fb>\u003C\u002Fa> ·\n\u003Ca href=\"#-downloads\">\u003Cb>Downloads\u003C\u002Fb>\u003C\u002Fa> ·\n\u003Ca href=\"#-licenses\">\u003Cb>Licenses\u003C\u002Fb>\u003C\u002Fa> ·\n\u003Ca href=\"https:\u002F\u002Fwww.audarai.com\">\u003Cb>Audar API\u003C\u002Fb>\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003C\u002Fdiv>\n\n---\n\n**Audar-ASR-V1** is a family of Arabic-first, generative speech-recognition models from\n[AudarAI](https:\u002F\u002Fwww.audarai.com). It recasts transcription as audio-conditioned next-token prediction\nover a unified text vocabulary — a language-model decoder rather than a CTC or transducer objective — and\nis adapted in-house through **300,000+ hours** of labeled Arabic audio and a four-stage curriculum that\nends in **KTO preference alignment** from native Arabic annotators. It transcribes MSA and every major\nArabic dialect (Gulf\u002FEmirati, Egyptian, Levantine, Maghrebi), code-switched Arabic–English, and English —\n**30 languages** in total.\n\nOn the **Open Universal Arabic ASR Leaderboard**, the Turbo tier posts the **lowest average WER and CER of\nany evaluated system (rank #1 of 36)**, and the 0.78 B Flash tier is the **strongest small model on the\nboard (#11)** — beating models several times its size. This repository is the developer hub: model\npointers, benchmarks, and copy-paste inference for both tiers.\n\n## 🧩 Models\n\n| | **Audar-ASR-V1-Flash** | **Audar-ASR-V1-Turbo** |\n|---|---|---|\n| **Tier** | Real-time · edge | Accuracy |\n| **Parameters** | 0.78 B (0.60 B decoder + 0.19 B encoder) | 2.35 B (2.03 B decoder + 0.32 B encoder) |\n| **Runtimes** | 🤗 Transformers · GGUF (llama.cpp) | GGUF (llama.cpp) |\n| **Leaderboard** | #11 of 36 · 33.31 % avg WER | **#1 of 36 · 24.78 % avg WER** |\n| **Best for** | Live captioning, voice agents, on-device \u002F offline | Lowest error on hard dialectal & long-form audio |\n| **License** | [AudarAI Open v1.0](https:\u002F\u002Fwww.audarai.com\u002Flicense\u002Faudarai-open-license-v1.0\u002F) | [AudarAI Community v1.0](https:\u002F\u002Fwww.audarai.com\u002Flicense\u002Faudarai-community-license-v1.0\u002F) |\n| **Download** | **[🤗 audarai\u002FAudar-ASR-V1-Flash](https:\u002F\u002Fhuggingface.co\u002Faudarai\u002FAudar-ASR-V1-Flash)** | **[🤗 audarai\u002FAudar-ASR-V1-Turbo](https:\u002F\u002Fhuggingface.co\u002Faudarai\u002FAudar-ASR-V1-Turbo)** |\n\nBoth share one architecture (Whisper-style 128-mel audio encoder → Qwen3 decoder, 30 s context) and one\nprompt interface, so you can develop against Flash and scale to Turbo without code changes.\n\n## ⚡ Quickstart\n\nWeights download automatically from the Hugging Face repos above.\n\n### Transformers — Flash (Python)\n\n```bash\npip install -r examples\u002Frequirements.txt\npython examples\u002Ftranscribe.py clip.wav            # Arabic (auto-dialect)\npython examples\u002Ftranscribe.py english.wav --lang en\n```\n\n```python\n# Or call the reference helpers directly:\nfrom audar_asr import load_model, transcribe\nmodel, proc = load_model(\"audarai\u002FAudar-ASR-V1-Flash\")   # HF repo id or local path\nprint(transcribe(model, proc, \"clip.wav\"))               # \u003C= 30 s clip\n```\n\n### GGUF — Flash or Turbo (llama.cpp, CPU\u002FGPU\u002Fedge)\n\n```bash\n# Build a recent llama.cpp with Qwen3-ASR support, then:\n.\u002Fexamples\u002Fgguf_infer.sh clip.wav turbo     # or: flash\n```\n\n```bash\n.\u002Fbuild\u002Fbin\u002Fllama-mtmd-cli \\\n  -m       Audar-ASR-V1-Turbo-Q8_0.gguf \\\n  --mmproj mmproj-Audar-ASR-V1-Turbo.gguf \\\n  --audio  clip.wav \\\n  -sys     \"فرّغ الكلام العربي التالي.\" \\\n  --temp 0\n```\n\n> ⚠️ The audio projector (`mmproj`) must stay **BF16** — its `ClippableLinear` is numerically sensitive.\n> The decoder GGUF quantizes normally (Q4_K_M \u002F Q8_0 \u002F BF16 all published on Hugging Face).\n\n### Long-form & realtime streaming\n\n```bash\npython examples\u002Ftranscribe_long.py meeting.wav   # arbitrary length (30 s chunking)\npython examples\u002Fstream.py long.wav               # LocalAgreement-2 incremental output\n```\n\nStreaming commits a word only once two consecutive sliding-window decodes agree on it, so committed text\nis stable and never rewrites. This is the local reference policy; Audar's production engine serves the\nsame policy with **sub-250 ms latency** over an OpenAI-Realtime-compatible WebSocket — see the\n[Audar API](https:\u002F\u002Fwww.audarai.com).\n\n## 📊 Benchmarks\n\nFull-test-set results on the **[Open Universal Arabic ASR Leaderboard](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002FUBC-NLP\u002Fopen-universal-arabic-asr-leaderboard)**,\nsame harness and normalizer as the public board (calibrated within 0.03 pp). Top of the table:\n\n| # | Model | Avg WER | Avg CER |\n|--:|---|--:|--:|\n| **1** | **Audar-ASR-V1-Turbo** | **24.78** | **9.49** |\n| 2 | CohereLabs\u002Fcohere-transcribe-arabic-07-2026 | 25.87 | 11.80 |\n| 3 | omnilingual-asr\u002FomniASR_LLM_7B | 28.32 | 12.52 |\n| … | | | |\n| **11** | **Audar-ASR-V1-Flash** (0.78 B) | **33.31** | **13.66** |\n| 12 | Qwen\u002FQwen3-ASR-1.7B | 33.36 | 12.33 |\n\n**Full 36-system standings, per-dataset WER\u002FCER, methodology, and machine-readable\n[`leaderboard.csv`](benchmarks\u002Fleaderboard.csv) → [benchmarks\u002F](benchmarks\u002FREADME.md).**\n\nHighlights: Turbo is **#1 overall** and best on dialect-heavy **SADA** and **MGB-2**; Flash (0.78 B) beats\n**Qwen3-ASR-1.7B**, Voxtral-Small-24B, and Whisper-large-v3. On in-house Gulf-Emirati long-form, Turbo hits\n**19.4 % WER \u002F 7.3 % CER**.\n\n## 📥 Downloads\n\nAll weights and GGUF variants live on Hugging Face:\n\n- **Flash** (Transformers + GGUF): **https:\u002F\u002Fhuggingface.co\u002Faudarai\u002FAudar-ASR-V1-Flash**\n- **Turbo** (GGUF): **https:\u002F\u002Fhuggingface.co\u002Faudarai\u002FAudar-ASR-V1-Turbo**\n- Full model collection: **https:\u002F\u002Fhuggingface.co\u002Faudarai**\n\nPrefer a managed, production-hosted endpoint (streaming, speaker-attributed transcription, diarization)?\nSee the **[Audar API](https:\u002F\u002Fwww.audarai.com)**.\n\n## 🌍 Languages & capabilities\n\n- **Primary**: Arabic — MSA and dialectal (Gulf\u002FEmirati, Egyptian, Levantine, Maghrebi), plus\n  **code-switched Arabic–English**; emits dialect-faithful orthography from audio alone.\n- **Also**: English + 28 additional languages (30 total).\n- **Tasks**: transcription (audio → UTF-8 text), prompt-steerable for language and formatting; long-form\n  and realtime streaming via LocalAgreement-2.\n- **Not** for covert speaker identification.\n\n## 📜 Licenses\n\n- **This repository** (example code, reference helpers, docs) — **Apache-2.0** ([LICENSE](LICENSE)).\n- **Model weights** — released under AudarAI model licenses (not Apache-2.0):\n  - Audar-ASR-V1-Flash → **[AudarAI Open License v1.0](https:\u002F\u002Fwww.audarai.com\u002Flicense\u002Faudarai-open-license-v1.0\u002F)** (commercial use, redistribution, and fine-tuning permitted).\n  - Audar-ASR-V1-Turbo → **[AudarAI Community License v1.0](https:\u002F\u002Fwww.audarai.com\u002Flicense\u002Faudarai-community-license-v1.0\u002F)** (research and limited commercial use for qualifying Community Entities).\n- **Enterprise \u002F large-scale \u002F model-as-a-service** use may require an **AudarAI Enterprise License** —\n  contact **contact@audarai.com** or visit **[audarai.com](https:\u002F\u002Fwww.audarai.com)**.\n\n## 📄 Citation\n\nTechnical report — *Audar-ASR-V1: A Multilingual, Arabic-First Generative Speech Recognition Foundation\nModel*, Audar AI Team, 2026 — **[read the PDF](report\u002FAudar-ASR-V1-Technical-Report.pdf)**.\n\n```bibtex\n@techreport{audar-asr-v1-2026,\n  title       = {Audar-ASR-V1: A Multilingual, Arabic-First Generative Speech Recognition Foundation Model},\n  author      = {Audar AI Team},\n  institution = {AudarAI},\n  year        = {2026},\n  url         = {https:\u002F\u002Fgithub.com\u002FAudarAI\u002FAudar-ASR-V1}\n}\n```\n\n---\n\n\u003Cdiv align=\"center\">\n\n### Leading Arabic-First Multilingual Audio Intelligence\n\n*AudarAI starts with Arabic — and expands to the world.*\n\n**Arabic-first. Multilingual by design. Human-centered at heart.**\n\n**[🌐 www.audarai.com](https:\u002F\u002Fwww.audarai.com)** · [🤗 Hugging Face](https:\u002F\u002Fhuggingface.co\u002Faudarai) · [GitHub](https:\u002F\u002Fgithub.com\u002FAudarAI) · contact@audarai.com\n\n© 2026 AUDARAI PTE. LTD.\n\n\u003C\u002Fdiv>\n","Audar-ASR-V1 是面向阿拉伯语的生成式语音识别模型家族，专为现代阿拉伯语（含标准语及海湾、埃及、黎凡特、马格里布等主要方言）和阿拉伯语-英语混杂语音设计。采用音频条件下的自回归文本生成范式（Whisper-style编码器+Qwen3解码器），支持GGUF格式与llama.cpp运行时，提供Flash（轻量实时）和Turbo（高精度）双版本，在Open Universal Arabic ASR排行榜中Turbo版位列第一（平均WER 24.78%）。适用于阿拉伯语语音转写、实时字幕、离线语音助手及多语种语音分析等场景。",2,"2026-07-13 02:30:05","CREATED_QUERY"]