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

const transcriber = await pipeline(
  'automatic-speech-recognition',
  'galamkhoahoc/EraX-WoW-Turbo-V1.1-ONNX',
  {
    device: 'wasm',
    dtype: {
      encoder_model: 'q8',
      decoder_model_merged: 'q8',
    },
  },
);

const result = await transcriber(audio16k, {
  language: 'vi',
  task: 'transcribe',
});

Model files are downloaded on first use and can be reused through the browser Cache API. Audio inference runs locally in the user's Web Worker and does not need to be sent to an application backend.

Export details

Transformers.js converter: 3.8.1
Transformers: 4.49.0
PyTorch exporter: 2.6.0
ONNX Runtime quantization: Q8
Audio input: mono 16 kHz

The generated ONNX graphs passed onnx.checker.check_model(..., full_check=True) before upload.