Modelpublic

dmouayad/whisper-large-v3-turbo-arabic-dialectal-v2-gguf

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes

Model Card

Whisper Large v3 Turbo — Arabic Dialectal V2 (GGUF)

GGML conversions and quantisations of oddadmix/whisper-large-v3-turbo-arabic-dialectal-v2 for use with CrispASR and whisper.cpp.

Available variants

File	Quant	Size	Notes
`ggml-model.bin`	F16	1.6 GB	Original conversion, full precision
`ggml-model-q5_0.bin`	Q5_0	574 MB	Good quality/size tradeoff
`ggml-model-q4_k.bin`	Q4_K	452 MB	Smallest, fastest on CPU

Usage with CrispASR

bash

# Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR && cd CrispASR
cmake -S . -B build && cmake --build build -j8
# Transcribe audio
./build/bin/crispasr -m ggml-model-q4_k.bin -f audio.wav

Conversion

Converted from the original HuggingFace model using whisper.cpp's convert-h5-to-ggml.py, then quantised with whisper-quantize:

bash

python CrispASR/models/convert-h5-to-ggml.py \
    ./whisper-large-v3-turbo-arabic-dialectal-v2 \
    ./whisper \
    ./output-ggml
whisper-quantize ggml-model.bin ggml-model-q5_0.bin q5_0
whisper-quantize ggml-model.bin ggml-model-q4_k.bin q4_k