flackzz/distil-whisper-large-v3-german_timestamped-ONNX

distil-whisper-large-v3-german_timestamped-ONNX

This repository contains ONNX weights for primeline/distil-whisper-large-v3-german prepared for use with Transformers.js.

Timestamp support is preserved through the exported Whisper generation config. Available dtypes in this export: fp32, fp16, q4, q8.

The repository layout follows the standard Transformers.js Whisper convention:

• processor, tokenizer, and config files in the repository root
• encoder and merged decoder ONNX files inside onnx/

Usage (Transformers.js)

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

const transcriber = await pipeline("automatic-speech-recognition", "<your-hf-username>/distil-whisper-large-v3-german_timestamped-ONNX", {
  device: "webgpu",
  dtype: "fp16", // or "fp32", "q4", "q8"
});

const output = await transcriber("audio.mp3", {
  chunk_length_s: 30,
  return_timestamps: "word",
});

console.log(output.text);
console.log(output.chunks?.slice(0, 3));

Recommended choices:

• fp32: highest precision, typically for server-side or high-memory WebGPU
• fp16: smaller WebGPU model with good speed/quality tradeoff
• q8: smaller CPU/WASM model
• q4: smallest model, best for constrained devices

Source Model

• Base model: primeline/distil-whisper-large-v3-german
• Export format: ONNX
• Intended runtime: Transformers.js