Modelpublic

elgehelge/plapre-onnx-web

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes

Model Card

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 or nano (see below); the Kanade files are shared.

File	Stage	Notes
`<variant>/lm/model.onnx` (+ `meta.json`)	Plapre language model	SmolLM2/LLaMA, autoregressive
`kanade_decoder.onnx` (+ `.onnx.data`)	Kanade decoder	audio tokens + speaker emb → mel
`hift_vocoder.onnx` (+ `.onnx.data`)	HiFT vocoder	mel → 24 kHz waveform
`clone_encoder.onnx`	Kanade clone encoder	reference audio → 128-dim speaker embedding
`<variant>/tokenizer.json`, `config.json`	tokenizer / config	BPE + special audio tokens
`<variant>/speakers.json`, `speaker_proj.json`	built-in voices	precomputed embeddings (`tor, ida, liv, ask, kaj`)

Model variants

The repo hosts both Plapre sizes. The LM-side files above are variant-specific and live under each variant's directory: Pico (hidden 576) under pico/ and Nano (hidden 960) under nano/ (e.g. pico/lm/model.onnx, 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" }).

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:

Artifact	Upstream	License
`<variant>/lm/*`, `tokenizer.json`, `config.json`, `speakers.json`, `speaker_proj.json`	Plapre Pico / Nano by syv.ai	CC BY 4.0
`kanade_decoder.*`, `clone_encoder.onnx`	Kanade (frothywater)	MIT
`hift_vocoder.*`	HiFT, from CosyVoice 2	Apache-2.0
clone-encoder SSL frontend	WavLM (Microsoft)	MIT

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