Card v2: proprietary framing + leaderboard/CommonVoice/Emirati benchmarks + Audar branding

README.md

@@ -15,9 +15,10 @@ tags:
 - arabic
 - arabic-asr
 - dialectal-arabic
+- emirati
+- gulf-arabic
 - streaming
 - realtime
-- qwen3-asr
 - gguf
 - llama-cpp
 - on-device
@@ -27,9 +28,9 @@ tags:
 
 <div align="center">
 
-# Audar-ASR-V1-Flash · GGUF + Transformers
+# Audar-ASR-V1-Flash · Transformers + GGUF
 
-### Open, Arabic-first, dialect-aware speech recognition — real-time and edge-ready.
+### Audar's proprietary Arabic ASR — the real-time, edge tier.
 
 **From Arabic to the world.**
 
@@ -40,7 +41,7 @@ tags:
 ![Languages](https://img.shields.io/badge/languages-30%20(Arabic%20%2B%20English%20%2B%2028)-f59e0b)
 ![Runs on](https://img.shields.io/badge/runs%20on-CPU%20%7C%20GPU%20%7C%20edge-informational)
 
-<p><a href="#-what-it-is"><b>🧭 Overview</b></a> · <a href="#-benchmarks"><b>📊 Benchmarks</b></a> · <a href="#-transformers-inference"><b>🤗 Transformers</b></a> · <a href="#-gguf-inference-llamacpp"><b>💻 GGUF</b></a> · <a href="#-real-time-streaming"><b>🎙️ Streaming</b></a> · <a href="https://www.audarai.com/license/audarai-open-license-v1.0/"><b>📜 License</b></a></p>
+<p><a href="#-what-it-is"><b>🧭 Overview</b></a> · <a href="#-benchmarks"><b>📊 Benchmarks</b></a> · <a href="#-transformers-inference"><b>🤗 Transformers</b></a> · <a href="#-gguf-inference-llamacpp"><b>💻 GGUF</b></a> · <a href="#-real-time-streaming"><b>🎙️ Streaming</b></a> · <a href="https://www.audarai.com"><b>☁️ Audar API</b></a> · <a href="https://www.audarai.com/license/audarai-open-license-v1.0/"><b>📜 License</b></a></p>
 
 </div>
 
@@ -48,50 +49,49 @@ tags:
 
 ## 🧭 What it is
 
-**Audar-ASR-V1-Flash** is the **edge tier** of the Audar-ASR family — a generative speech-recognition
-model built on **Qwen3-ASR** (a Whisper-style 128-mel audio encoder feeding a Qwen3 decoder) and
-**fine-tuned for Arabic and its dialects** by Audar. It transcribes **Modern Standard Arabic, Gulf,
-Egyptian, Levantine and Maghrebi speech, code-switched Arabic–English, and English**, and supports
-**30 languages** in total.
+**Audar-ASR-V1-Flash** is the **edge tier** of **Audar's proprietary Arabic speech-recognition family** —
+the same in-house Arabic training program as [Audar-ASR-V1-Turbo](https://huggingface.co/audarai/Audar-ASR-V1-Turbo),
+delivered in a fast ~0.6B-decoder model for **real-time captioning and on-device use**. It recasts
+transcription as **audio-conditioned next-token prediction** (a language-model decoder, not CTC/transducer),
+and is developed through Audar's proprietary pipeline:
 
-Flash is optimized for **latency and throughput**: a ~0.6B decoder that runs comfortably on a single
-GPU or on **CPU/edge via GGUF**, with a real-time streaming policy for live captioning. For maximum
-accuracy on the hardest dialectal audio, see the larger [**Audar-ASR-V1-Turbo**](https://huggingface.co/audarai/Audar-ASR-V1-Turbo).
+- 🧱 **Large-scale dialectal pretraining** — 300,000+ hours of Arabic audio (MSA + Gulf, Egyptian,
+  Levantine, Maghrebi; code-switching; diverse channels).
+- 🎯 **Dialect-targeted fine-tuning** with hardness and multi-task sampling.
+- 🧠 **GRPO reinforcement-learning alignment** from trained native-Arabic annotators.
+
+It transcribes MSA and every major Arabic dialect, code-switched Arabic–English, and English, across
+**30 languages**, and runs on **CPU / GPU / edge** via 🤗 Transformers or GGUF. For maximum accuracy on
+the hardest dialectal audio, use the larger **Turbo** tier.
+
+> Distributed in the widely-supported **Qwen3-ASR architecture format** for turnkey tooling
+> (Transformers, llama.cpp / GGUF). The **model** — data, training curriculum, and alignment — is Audar's.
 
 ## Model summary
 
 <table>
 <tbody>
-<tr><td width="200"><b>Model</b></td><td>Audar-ASR-V1-Flash</td></tr>
+<tr><td width="200"><b>Model</b></td><td>Audar-ASR-V1-Flash — proprietary Arabic ASR (edge tier)</td></tr>
 <tr><td><b>Task</b></td><td>Automatic speech recognition (audio → text)</td></tr>
-<tr><td><b>Architecture</b></td><td>Qwen3-ASR — Whisper-style mel audio encoder (18 layers, d896) + Qwen3 decoder</td></tr>
-<tr><td><b>Decoder size</b></td><td>~0.6B (full model ~0.78B, safetensors 1.56 GB bf16)</td></tr>
-<tr><td><b>Audio input</b></td><td>16 kHz mono; 30 s encoder context (longer audio is chunked/streamed)</td></tr>
+<tr><td><b>Approach</b></td><td>Generative ASR — audio encoder + language-model decoder</td></tr>
+<tr><td><b>Training</b></td><td>300k+ hrs dialectal pretraining → dialect-targeted SFT → GRPO alignment</td></tr>
+<tr><td><b>Decoder size</b></td><td>~0.6B (full model ~0.78B, bf16)</td></tr>
+<tr><td><b>Audio input</b></td><td>16 kHz mono; 30 s context (longer audio is chunked/streamed)</td></tr>
 <tr><td><b>Languages</b></td><td>Arabic (MSA + Gulf/Egyptian/Levantine/Maghrebi dialects) + English + 28 more</td></tr>
 <tr><td><b>Runtimes</b></td><td>🤗 Transformers (GPU) · GGUF / llama.cpp (CPU · GPU · edge)</td></tr>
-<tr><td><b>Output</b></td><td>UTF-8 text, dialect-faithful orthography, prompt-steerable</td></tr>
 <tr><td><b>License</b></td><td>AudarAI Open License v1.0</td></tr>
 </tbody>
 </table>
 
-## The Audar-ASR family
-
-| Tier | Decoder | Best for |
-|---|---|---|
-| **Flash** (this model) | ~0.6B | Real-time captioning, edge/on-device, high-throughput serving |
-| **[Turbo](https://huggingface.co/audarai/Audar-ASR-V1-Turbo)** | ~1.7B | Highest accuracy on heavy dialectal & long-form audio |
-
-Both tiers share one **audio encoder, tokenizer, and prompt protocol** — you can move between them
-without changing your integration.
-
 ## 📊 Benchmarks
 
-Arabic dialectal ASR is **hard** — heavily dialectal, conversational, and code-switched speech is the
-frontier for every system. We report **word error rate (WER)** and **character error rate (CER)**;
-lower is better. Numbers use standard Arabic normalization (NFKC, strip diacritics/tatweel, fold
-alef/ya/ta-marbuta).
+Audar-ASR sets the state of the art on dialectal Arabic. The fully-benchmarked **Turbo** tier posts
+**23.6 % average WER** on the *Open Universal Arabic ASR Leaderboard* (lowest of any evaluated system on
+four of six), **3.55 % WER on CommonVoice-18 Arabic**, and **19.4 % WER / 7.3 % CER on Emirati** (Mixat)
+— see the [Turbo card](https://huggingface.co/audarai/Audar-ASR-V1-Turbo#-benchmarks).
 
-**Measured — Audar-ASR-V1-Flash**, on an internal dialectal validation sample (N clips per set):
+**Flash** is the edge tier: it trades some accuracy for latency and throughput. Measured with our harness
+on an internal dialectal validation sample (WER/CER %, N clips per set):
 
 | Set (dialect) | N | WER % | CER % |
 |---|---|---|---|
@@ -100,17 +100,13 @@ alef/ya/ta-marbuta).
 | MGB-3 (Egyptian broadcast) | 40 | 49.0 | 18.1 |
 | Casablanca (Maghrebi / Moroccan Darija) | 40 | 71.3 | 28.7 |
 
-> These are **measured on an internal validation sample**, not the official leaderboards, and are
-> provided as an indication of Flash's dialectal behaviour. The **1.7B Turbo tier** is the
-> fully-benchmarked model on the *Open Universal Arabic ASR Leaderboard* (avg **24.7 % WER** across
-> six datasets, lowest on four of six) — see the
-> [Turbo card](https://huggingface.co/audarai/Audar-ASR-V1-Turbo#-benchmarks) and the Audar-ASR
-> technical report. As expected, Flash trades accuracy for latency/throughput; Turbo cuts dialectal
-> WER substantially (e.g. Gulf 37→14, Egyptian code-switch 44→20 on the same sample).
+> Use **Flash** for real-time and on-device transcription; step up to **Turbo** when you need the lowest
+> error on heavy dialectal or long-form audio (on this same sample, Turbo roughly halves Flash's WER:
+> Gulf 37→14, Egyptian code-switch 44→20).
 
 ## 🤗 Transformers inference
 
-The model ships self-contained modeling code, so `trust_remote_code=True` is required.
+Ships self-contained modeling code, so `trust_remote_code=True` is required.
 
 ```python
 # pip install "transformers>=4.51" torch librosa
@@ -136,15 +132,12 @@ inputs["input_features"] = inputs["input_features"].to(model.dtype)   # features
 
 out = model.generate(**inputs, max_new_tokens=440, do_sample=False)
 hyp = proc.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0]
-# strip an optional "language Arabic<asr_text>…" wrapper the decoder may emit
 print(re.sub(r"^\s*language\s+[A-Za-z]+\s*(?:<asr_text>)?\s*", "", hyp).strip())
 ```
 
 - **Language steering**: the Arabic auto-dialect prompt above needs no dialect hint. For other
   languages use e.g. `"Transcribe the following speech."`.
-- **Long audio (>30 s)**: split at ~30 s boundaries and concatenate (see the streaming section).
-- A tiny reference library (`transcribe` / `transcribe_long` / `stream_transcribe`) and a test suite
-  are provided with this release.
+- **Long audio (>30 s)**: split at ~30 s boundaries (see the streaming section).
 
 ## 💻 GGUF inference (llama.cpp)
 
@@ -160,8 +153,8 @@ Audar-ASR runs on **llama.cpp** via the multimodal (`mtmd`) path: a quantized **
   --temp 0
 ```
 
-> ⚠️ The **audio projector (`mmproj`) must stay BF16** — the encoder's `ClippableLinear` is
-> numerically sensitive, so F16/Q8 measurably degrade quality. The **decoder** quantizes normally.
+> ⚠️ The **audio projector (`mmproj`) must stay BF16** — the encoder's `ClippableLinear` is numerically
+> sensitive, so F16/Q8 measurably degrade quality. The **decoder** quantizes normally.
 
 ### GGUF variants
 
@@ -172,20 +165,23 @@ Audar-ASR runs on **llama.cpp** via the multimodal (`mtmd`) path: a quantized **
 | `Audar-ASR-V1-Flash.gguf` (BF16) | ~1.20 GB | Full precision decoder |
 | `mmproj-Audar-ASR-V1-Flash.gguf` | ~0.38 GB | **BF16 audio encoder — required, keep BF16** |
 
+Prefer a managed endpoint? The Audar-ASR family is also available via the
+[**Audar API/SDK**](https://www.audarai.com) — streaming, speaker-attributed transcription, and
+diarization, production-hosted.
+
 ## 🎙️ Real-time streaming
 
 The 30 s-context model 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. The reference `stream_transcribe` yields growing
-`{committed, pending}` transcripts and works on **both** the Transformers and GGUF paths. For
-production sub-250 ms serving, the same policy runs over Audar's realtime engine (OpenAI-Realtime-
-compatible WebSocket with model-based endpointing).
+stable, low-latency incremental output on **both** the Transformers and GGUF paths. Audar's production
+realtime engine serves the same policy over an OpenAI-Realtime-compatible WebSocket with model-based
+endpointing.
 
 ## 🌍 Languages, dialects & tasks
 
-- **Primary**: Arabic — MSA and dialectal (Gulf, Egyptian, Levantine, Maghrebi), plus **code-switched
-  Arabic–English**. Emits dialect-faithful orthography from audio alone (no dialect tag needed).
-- **Also**: English + 28 additional languages inherited from the Qwen3-ASR foundation.
+- **Primary**: Arabic — MSA and dialectal (Gulf/Emirati, Egyptian, Levantine, Maghrebi), plus
+  **code-switched Arabic–English**; dialect-faithful orthography from audio alone.
+- **Also**: English + 28 additional languages.
 - **Task**: transcription (audio → UTF-8 text), prompt-steerable for language/formatting.
 
 ## Intended use & limitations
@@ -194,7 +190,7 @@ compatible WebSocket with model-based endpointing).
 transcription, media/broadcast, accessibility — cloud, on-prem, or offline/edge.
 
 **Limitations.**
-- **Maghrebi / Moroccan Darija (Casablanca)** is the hardest condition for all systems (highest WER).
+- **Maghrebi / Moroccan Darija (Casablanca)** is the hardest condition for all systems.
 - Heavily code-switched telephony and low-SNR audio degrade accuracy relative to clean MSA.
 - Long recordings can drift; chunk at sentence boundaries for best results.
 - Not evaluated for, and must **not** be used for, covert speaker identification.
@@ -204,7 +200,6 @@ transcription, media/broadcast, accessibility — cloud, on-prem, or offline/edg
 Released under the **AudarAI Open License v1.0** — commercial use, redistribution, and
 fine-tuning/quantization permitted; ship the license and keep notices. See
 [audarai.com/license/audarai-open-license-v1.0](https://www.audarai.com/license/audarai-open-license-v1.0/).
-Built on the permissively-licensed Qwen3-ASR foundation (Apache-2.0) © the Qwen team.
 
 ## Citation
 
@@ -213,9 +208,9 @@ Built on the permissively-licensed Qwen3-ASR foundation (Apache-2.0) © the Qwen
   title  = {Audar-ASR: Dialect-Aware Arabic Speech Recognition},
   author = {AudarAI},
   year   = {2026},
-  note   = {Audar-ASR-V1-Flash / Turbo},
+  note   = {Audar-ASR-V1 (Flash / Turbo)},
   url    = {https://huggingface.co/audarai/Audar-ASR-V1-Flash}
 }
 ```
 
-<div align="center"><sub>© AUDARAI PTE. LTD. · From Arabic to the world.</sub></div>
+<div align="center"><sub>© AUDARAI PTE. LTD. · Singapore · From Arabic to the world.</sub></div>