Whisper base (Quran), for whisper.cpp
tarteel-ai/whisper-base-ar-quran, converted to the ggml format of whisper.cpp and quantized to 5 bits (q5_1). The weights are Tarteel's; only their format changed.
Used by Thumn, an Android app for memorizing the Quran, to follow a recitation from memory on the phone and mark the words missed or said wrong. Nothing leaves the phone.
| File | Size | SHA-256 |
|---|---|---|
ggml-base-quran-q5_1.bin |
59,707,625 bytes | acc1f03c8b1468b674b4fab50e0b2312cff982ba0001cd16fe36ca23862b6536 |
How it was made
With whisper.cpp v1.9.5:
python3 whisper.cpp/models/convert-h5-to-ggml.py whisper-base-ar-quran openai-whisper .
whisper-quantize ggml-model.bin ggml-base-quran-q5_1.bin q5_1
Tarteel's config.json has max_length: 1024, which the conversion script takes as the decoder's context;
it was set to 448 (max_target_positions) first, or whisper.cpp refuses the model.
License
Apache 2.0, as the original model. Credit: Tarteel AI, and OpenAI for Whisper.