whisper-atc-small-onnx

whisper-small.en ATC (ATCO2) — ONNX for transformers.js

ONNX conversion of jlvdoorn/whisper-small.en-atco2-asr (TU Delft Whisper-ATC, fine-tuned on the ATCO2 real-radio corpus), laid out for transformers.js / onnxruntime-web.

~3x faster than the medium variant (Japb62/whisper-atc-medium-onnx) — intended for near-real-time per-transmission transcription in the browser.

Files

onnx/encoder_model_fp16.onnx               (~168 MB)  <- use with WebGPU
onnx/decoder_model_merged_fp16.onnx        (~370 MB)  <- use with WebGPU
onnx/encoder_model_quantized.onnx          (~88 MB)   <- use with WASM (q8)
onnx/decoder_model_merged_quantized.onnx   (~187 MB)  <- use with WASM (q8)
config.json, generation_config.json, preprocessor_config.json, tokenizer files

Usage (transformers.js)

import { pipeline } from '@huggingface/transformers';

const asr = await pipeline(
  'automatic-speech-recognition',
  'Japb62/whisper-atc-small-onnx',
  {
    device: 'webgpu', // or 'wasm'
    dtype: { encoder_model: 'fp16', decoder_model_merged: 'fp16' }, // wasm: 'q8'
  }
);
// English-only model: do NOT pass `language`/`task`.
const out = await asr(audio);

Conversion

Same recipe as the medium repo: optimum 1.23.3 + transformers 4.42.4 (legacy TorchScript exporter, opset 18, transformers < 4.43 to avoid the cache_position input that transformers.js does not feed); q8 = uint8 dynamic quantization; fp16 = native CUDA export with graph IO cast back to fp32.