QwenCleo-ASR

🚀 Quick start

pip install qwencleo-asr

from qwencleo_asr import QwenCleoASR

asr = QwenCleoASR()                       # loads mohammedaly22/QwenCleo-ASR
print(asr.transcribe("clip.wav").text)

Or with the upstream qwen-asr API directly:

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "mohammedaly22/QwenCleo-ASR", dtype=torch.bfloat16, device_map="cuda:0",
)
result = model.transcribe(audio="clip.wav", language="Arabic")
print(result[0].text)

⚡ Streaming & serving (vLLM)

QwenCleo inherits Qwen3-ASR's true token-by-token streaming via vLLM (needs an Ampere-or-newer GPU: L4 / A100 / H100).

pip install qwencleo-asr
qwencleo install-vllm     # vLLM nightly (cu129) — only build with Qwen3-ASR support
qwencleo serve            # OpenAI-compatible server on :8000

from qwencleo_asr import QwenCleoASR

asr = QwenCleoASR()
for delta in asr.stream("clip.wav", port=8000):   # real streaming via vLLM
    print(delta, end="", flush=True)

OpenAI-compatible transcription, a live mic web demo (qwen-asr-demo-streaming), offline VLLMOffline, and cURL examples are in the vLLM serving guide.

For captioning long files without a server, stream_file(...) does chunked (windowed) transcription — convenient, but latency is per-window, not per-token.

🗣️ Sample outputs

Real transcriptions from the test set, scored after Egyptian-aware normalization. Examples span short, medium, and long clips in both Egyptian Arabic and code-switching. These are picked to be honest — on some clips other models also do well; QwenCleo's edge is consistency, dialect fidelity, and keeping English terms in Latin script.

✅ = matches ground truth · ⚠️ = minor slips · ❌ = clear errors. Bold marks the wrong spans.

🔀 Code-switching

Short

✅ Ground truth ااه بحس ستايله حلو وشكله حلو كاريزما وهو بيلعب

Medium

✅ Ground truth — ده حصل ازاي قرار اصلا انك تعمل قناة YouTube

✅ Ground truth — خليني اوضح بس هدفها ان الطالب يجي هو already

Long

✅ Ground truth — ما هو الفكرة جاية فين النهارده average العربية اللي ممكن تجيبها من أوروبا ما بين 23 ل 28000 dollars

✅ Ground truth — فبتديك كل الsoft skills بشكل indirect حرفيا بتديهالك بالمعلقه ليه بتتعلم تشتغل على كل الحاجات بتاعه Microsoft Excel PowerPoint Word whatever بعدين كمان بتتعلم ان انت تتكلم مع الناس public speaking

🇪🇬 Egyptian Arabic

Pure-Arabic clips (no English). After normalization these reflect genuine word-level accuracy, not spelling/number-format differences.

Short

✅ Ground truth — الشهرة مرت بمرحلتين معاك

Medium

✅ Ground truth — لا ركز عشان انت دلوقتي هتتحاسب على الكلام ده

✅ Ground truth — بصراحة كانت من أسعد أيام حياتي لما شفت اللي هما فرحانين

Long

✅ Ground truth — انا النهاردة مثلا ساكن في مدينة نصر وشغلي في المعادي فانا من البيت للشغل بقطع 20 كيلو في اليوم

✅ Ground truth — اشتغلت كان فيه معاناة في تحضير المنهج

🧠 Model details

• Base model: Qwen/Qwen3-ASR-1.7B (repo)
• Fine-tuning data: hundreds of hours of Egyptian podcast speech (Egyptian Arabic + Arabic/English code-switching).
• Languages: Egyptian Arabic, Modern Standard Arabic, Arabic ↔ English code-switching.
• Recommended language hint: "Arabic" (or None to auto-detect).
• Intended use: transcription/captioning of Egyptian Arabic and code-switched speech.
• Limitations: tuned for Egyptian dialect; other Arabic dialects and noisy far-field audio may degrade. Not optimized for pure-English audio.

🔗 Links

• 💻 GitHub: MohammedAly22/qwencleo-asr — package source, server, Gradio app, vLLM streaming guide.
• 📦 PyPI: qwencleo-asr
• 🧱 Base: Qwen3-ASR-1.7B · Qwen3-ASR repo

📜 Citation

@misc{qwencleo_asr_2026,
  title  = {QwenCleo-ASR: The Best Open-Source Egyptian Arabic and Code-Switching Speech Recognition Model},
  author = {Mohammed Aly},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/mohammedaly22/QwenCleo-ASR}},
  note   = {Fine-tuned from Qwen3-ASR-1.7B}
}

License: Apache-2.0, inheriting the Qwen3-ASR license terms.