plapre-onnx-web

Plapre — ONNX weights for in-browser inference

Converted ONNX artifacts for running Plapre, an open-source Danish text-to-speech model, fully in the browser on onnxruntime-web (WebGPU, with a WASM fallback).

These files are consumed by the plapre-in-a-browser library and its live demo. They are a format conversion (PyTorch → ONNX) of the upstream model — no retraining — verified to reproduce the PyTorch reference to float precision.

Files

<variant> is pico, nano, v2-nano, or v3-nano (see below); the Kanade files are shared.

Model variants

The repo hosts every supported Plapre variant. The LM-side files above are variant-specific and live under each variant's directory: Plapre 1 Pico (hidden 576) under pico/, Plapre 1 Nano (hidden 960) under nano/, Plapre 2 Nano (hidden 960, retrained) under v2-nano/, and Plapre 3 Nano (hidden 960, RL-tuned upstream for clearly better intelligibility) under v3-nano/ (e.g. pico/lm/model.onnx, v3-nano/lm/model.onnx). The Kanade kanade_decoder.*, hift_vocoder.*, and clone_encoder.onnx are shared across variants and stay at the root. Select a variant with loadPlapreEngine({ model: "pico" | "nano" | "v2-nano" | "v3-nano" }).

Usage

import { loadPlapreEngine } from "plapre-in-a-browser";

const engine = await loadPlapreEngine({
  modelsBaseUrl: "https://huggingface.co/elgehelge/plapre-onnx-web/resolve/main",
});
const { samples, sampleRate } = await engine.synthesizeToPcm({
  text: "Hej, hvordan har du det i dag?",
  voice: "ida",
});

License & attribution

Distributed under CC BY 4.0 (the most restrictive of the upstream licenses). Each artifact derives from a permissively licensed work:

These are format conversions (PyTorch → ONNX), not retraining. If you redistribute these artifacts, keep this attribution and the MIT/Apache notices. Full notices: https://github.com/elgehelge/plapre-in-a-browser/blob/main/NOTICE