qroudry/clipsun-models

Clipsun AI model files

Model files downloaded by the Clipsun AI desktop app on first run. They are format conversions of biodatlab/whisper-th-large-v3-combined (Thonburian Whisper, Thai fine-tune of OpenAI Whisper large-v3), licensed Apache-2.0. The weights are unchanged apart from the format and float16 precision.

Folder	Format	Used on
whisper-th-large-v3-combined/	CTranslate2, float16 (for faster-whisper)	NVIDIA graphics cards (CUDA)
whisper-th-large-v3-combined-onnx/	ONNX, float16, fixed-shape encoder + decoder graphs for ONNX Runtime DirectML	AMD / Intel graphics cards

The ONNX decoder takes one token per beam for one 15-second audio chunk (5 beams, 256 token slots) with per-layer key/value caches, and also returns the cross-attention of the alignment heads (align_heads.json) for word timestamps.

Credits

• Thonburian Whisper by Biomedical and Data Lab, Mahidol University (biodatlab), Apache-2.0.
• OpenAI Whisper, MIT.

See LICENSE (Apache License 2.0).