ONNX conversion of typhoon-ai/typhoon-whisper-turbo — Typhoon Whisper Turbo (SCB 10X / Typhoon) — for
Transformers.js, exported with
cross-attention outputs so return_timestamps: 'word' works in the browser.
All credit for the model goes to the original authors. License: MIT (same as the base model).
| File | Use |
|---|---|
onnx/encoder_model_fp16.onnx, onnx/decoder_model_merged_fp16.onnx |
WebGPU (dtype: 'fp16') |
onnx/encoder_model_quantized.onnx, onnx/decoder_model_merged_quantized.onnx |
WASM / CPU (dtype: 'q8') |
scripts/convert.py --output_attentions
(transformers 4.49.0, optimum as pinned there, torch 2.6.0), task
automatic-speech-recognition-with-past.scripts/quantize.py (fp16, q8 per-channel).DequantizeLinear of its uint8 copy (same predictions, much smaller file).import { pipeline } from '@huggingface/transformers';
const asr = await pipeline('automatic-speech-recognition', 'supersuphot/typhoon-whisper-turbo-timestamped', {
dtype: { encoder_model: 'q8', decoder_model_merged: 'q8' },
});
const out = await asr(audio16kMono, { language: 'thai', task: 'transcribe', return_timestamps: 'word' });
onnx/encoder_model_fp16.onnx 1.3 GBonnx/encoder_model_quantized.onnx 647 MBonnx/decoder_model_merged_fp16.onnx 344 MBonnx/decoder_model_merged_quantized.onnx 173 MB