typhoon-whisper-large-v3, ONNX with word timing

ONNX conversion of typhoon-ai/typhoon-whisper-large-v3 (MIT) for running in the browser with transformers.js. The decoder also outputs its cross-attention maps, which word-level timing is computed from.

Changes from the original files:

• Weights converted to ONNX in 4-bit (q4f16, for WebGPU) and 8-bit (quantized, and uint8 for processors without VNNI) variants. No retraining.
• alignment_heads are the original Whisper large-v3 ones. In the browser their word start times agreed with those of the separately checked typhoon-whisper-turbo conversion to a median of 20 ms (90% within 60 ms) on 140 s of Thai talk-show speech.

The model was trained without timestamp tokens: generate with return_timestamps: false and return_token_timestamps: true.

Measured on 60 TVSpeech clips (Thai talk shows with English words), PyTorch half precision: 5.75% character error rate, against 6.79% for typhoon-whisper-turbo.

All credit for the model goes to its original authors (typhoon-ai).