reomar15/QwenCleo-ASR-transcribe-cpp-GGUF

QwenCleo-ASR for transcribe.cpp / Hex

Q8_0 conversion of Mohammed Aly's QwenCleo-ASR, a Qwen3-ASR 1.7B fine-tune for Egyptian Arabic and Arabic/English code-switching. Base model: Alibaba Qwen3-ASR 1.7B. The model and base model are declared Apache-2.0; see LICENSE. This packaging is for transcribe.cpp, not the author's audio.cpp GGUF runtime.

Provenance

• Source checkpoint: mohammedaly22/QwenCleo-ASR@50f6b901a3b882bdf28e71efd91f9f561510fc6f.
• transcribe.cpp converter and quantizer: 7045d1c0fe7cee37747b2792977421d435e75fca.
• File: QwenCleo-ASR-1.7B-Q8_0.gguf; bytes: 2185030816.
• SHA-256: 43bae9d5d7a26684b4a0faf1abc6e872b5825cc9147dbb7200da9ff4c5d12f87.
• Architecture: qwen3_asr; variant: qwen3-asr-1.7b.
• Tensor types: F32, Q8_0.
• Recipe: Hex fork, scripts/convert-qwencleo.py, then transcribe-quantize --quant Q8_0. The wrapper checks the pinned source weight hash and tied embeddings and changes attribution metadata; the backend performs tensor conversion.

Usage and limits

Use transcribe.cpp with 16 kHz mono audio and an explicit ar language hint:

transcribe-cli -m QwenCleo-ASR-1.7B-Q8_0.gguf --language ar speech.wav

The native backend supports offline transcription, cancellation and automatic long-audio chunking in Hex. It does not provide word timestamps or streaming. The Hex catalog exposes Arabic only: mixed English terms are the intended use, not a claim of general English or inherited 30-language quality.

Structural checks and native fixture inference are preparation checks, not an Egyptian speech benchmark, physical microphone test or installed Hex acceptance. No independent Egyptian WER or speed is claimed for this converted artifact. The source author's reported benchmark belongs to that source checkpoint and evaluation setup; do not treat it as a measurement of this Q8 conversion.