whisper-atc-medium-onnx

whisper-medium.en ATC — ONNX (q8) for transformers.js

ONNX q8 (uint8 dynamic) conversion of jacktol/whisper-medium.en-fine-tuned-for-ATC, laid out for transformers.js / onnxruntime-web (WebGPU + WASM fallback).

Fine-tuned for air traffic control (ATC) radio speech (English, ICAO phraseology). Intended for in-browser, offline transcription.

Files

onnx/encoder_model_fp16.onnx               (~586 MB)  <- use with WebGPU
onnx/decoder_model_merged_fp16.onnx        (~974 MB)  <- use with WebGPU
onnx/encoder_model_quantized.onnx          (~299 MB)  <- use with WASM (q8)
onnx/decoder_model_merged_quantized.onnx   (~491 MB)  <- use with WASM (q8)
config.json, generation_config.json, preprocessor_config.json,
tokenizer.json, tokenizer_config.json, vocab.json, merges.txt,
special_tokens_map.json, added_tokens.json, normalizer.json

Usage (transformers.js)

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

// WebGPU: use fp16. q8 (MatMulInteger) misbehaves on the WebGPU EP
// (garbage output, very slow) — q8 is for the WASM/CPU backend only.
const asr = await pipeline(
  'automatic-speech-recognition',
  'Japb62/whisper-atc-medium-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, { chunk_length_s: 30, return_timestamps: true });

Conversion

Exported with optimum 1.23.3 + transformers 4.42.4 (legacy TorchScript ONNX exporter, opset 18). q8 = uint8 dynamic quantization (onnxruntime). fp16 = native half-precision export on CUDA (fp16=True, device="cuda"), with graph inputs/outputs cast back to fp32 for transformers.js compatibility. Notes:

• transformers < 4.43 is required at export time: newer versions make optimum add a cache_position decoder input that transformers.js does not feed (runtime error "Missing the following inputs: cache_position").
• Post-hoc fp32→fp16 conversion (onnxconverter-common / onnxruntime float16) breaks the merged decoder's If subgraphs; native fp16 export is the way.
• q4 is intentionally avoided — it corrupts the Whisper decoder.