QwenCleo-ASR GGUF

Egyptian Arabic & code-switching speech recognition for audio.cpp: 3–4× faster than PyTorch on the same GPU, no Python

These are standalone GGUF packages of QwenCleo-ASR, a Qwen3-ASR-1.7B fine-tune for Egyptian Arabic and Arabic↔English code-switching (فبتديك كل ال soft skills بشكل indirect). On the author's 3,699-utterance Egyptian test set it reaches 19.85% WER, against 41.51% for base Qwen3-ASR-1.7B. Full results are on the original model card.

In audio.cpp it runs through the existing qwen3_asr family, the same C++ runtime and options as Qwen3-ASR: offline and streaming modes, long-audio chunking, word timestamps with Qwen3 Forced Aligner, and the OpenAI-compatible server.

📦 Files

Each file embeds the tokenizer, processor, config files and the audio.cpp model spec, so one file is all you need.

🚀 Usage

Build audio.cpp (Linux · Windows · macOS), then:

# install through the audio.cpp model manager (or download the .gguf directly)
audiocpp_model_manager install qwencleo_asr_q8_0 --models-dir models

audiocpp_cli --task asr --family qwen3_asr --model models/QwenCleo-ASR-GGUF --backend cuda \
  --audio speech.wav --language Arabic --text "" --text-out transcript.txt

• Pass --language Arabic. Quantized packages are less reliable at automatic language detection.
• Any sample rate works, because audio is resampled internally. Long recordings are chunked automatically.
• Streaming: --mode streaming --request-option audio_chunk_seconds=5.

OpenAI-compatible server:

{ "models": [{ "id": "qwencleo", "family": "qwen3_asr",
  "path": "models/QwenCleo-ASR-GGUF/qwencleo-asr-q8_0.gguf", "task": "asr" }] }

curl http://127.0.0.1:8080/v1/audio/transcriptions -F model=qwencleo -F language=Arabic -F file=@speech.wav

⚡ Speed: original PyTorch vs. audio.cpp

Same model, same GPU, same 16 Egyptian Arabic clips (75 s of audio: short lines, questions, code-switching, numbers, and a 22 s paragraph). Each runtime first loaded the model and ran one warm-up, then timed every clip.

• PyTorch: the official qwencleo-asr package (transformers backend), torch 2.11 + CUDA 12.6.
• audio.cpp: a long-lived audiocpp_server session.
• Hardware: GTX 1660 Ti Laptop 6 GB (Turing, no tensor cores), WSL2.

RTF is processing time divided by audio length; lower is faster.

• Cold start on this machine, with both runtimes reading from the same Windows drive: audio.cpp was ready in ~40 s, while PyTorch took 12–18 min to load the 4 GB safetensors in a 7.6 GB WSL VM.
• Speed and memory stay flat across a long-lived session: 33 consecutive requests ran at the same speed with identical output.
• ¹ -DCMAKE_CUDA_ARCHITECTURES="61-virtual;80-virtual" -DGGML_CUDA_FORCE_MMQ=ON. This is what ggml recommends for GTX 16xx cards. On GPUs with tensor cores, both runtimes are faster.

🎯 Output parity with PyTorch

These are transcripts of the same 16 clips, scored against the spoken text with Egyptian-aware normalization (diacritics and أ/إ/آ, ى/ي, ة/ه spelling variants ignored):

audio.cpp F16 and PyTorch bf16 give identical transcripts on 10 of 16 clips. The other six differ by one or two words at near-tied tokens, which is normal floating-point drift between runtimes, and the differences go both ways. For example, audio.cpp gets بيتعبني right where PyTorch wrote بيعبدي, while PyTorch is right on ملوخية. A 16-clip set is small, so treat differences under ~2 points as noise.

⚠️ Limitations

• Tuned for Egyptian Arabic and code-switching. Other dialects and noisy far-field audio may degrade.
• Numbers are usually written as digits (3:30, 20), as in the training transcripts.

📜 License & credits

• QwenCleo-ASR: Apache-2.0, by Mohammed Aly. Base model: Qwen/Qwen3-ASR-1.7B (Apache-2.0).
• Runtime: audio.cpp by @0xShug0. Conversion: audiocpp_gguf --type q8_0|f16 --family qwen3_asr.

@misc{qwencleo_asr_2026,
  title  = {QwenCleo-ASR: Egyptian Arabic and Code-Switching Speech Recognition},
  author = {Mohammed Aly},
  year   = {2026},
  url    = {https://huggingface.co/mohammedaly22/QwenCleo-ASR}
}