---
license: openrail
base_model:
- supertone-oss-archive/supertonic-2
language:
- ko
- en
- es
- pt
- fr
pipeline_tag: text-to-speech
library_name: onnx
tags:
- onnxruntime-web
- webgpu
---

# Supertonic 2 — web (fp16 weights)

Supertonic 2를 브라우저(onnxruntime-web, WebGPU·WASM)에서 가볍게 받도록 다시 저장한 것입니다.
서가(Seoga) 리더의 "서가 음성"이 이 파일을 받아 기기에서 합성합니다.

A browser-friendly repackaging of [Supertone Supertonic 2](https://huggingface.co/supertone-oss-archive/supertonic-2)
(revision `7edbcd0bba8cd579100a353a9e0b2b49cad7c875`) for onnxruntime-web (WebGPU / WASM).

## What changed

- Large float32 initializers (≥1024 elements) are stored as float16 and cast back to float32 by a `Cast` node at the
  start of each graph. ONNX Runtime folds these casts when the session is created, so **inference still runs in
  float32**; only the download size is halved (~263 MB → ~130 MB).
- Graph inputs/outputs, `tts.json`, `unicode_indexer.json` and the ten preset voice styles (`voices/F1–F5, M1–M5`) are
  unchanged.
- `webgpu/`: the same four models converted to **float16 compute** (inputs/outputs stay float32) for WebGPU.
  They use half the GPU memory because weights are not expanded to float32 — important on phones (iOS).
  Seoga uses them first on WebGPU and falls back to the root files (WebGPU without shader-f16, or WASM).
- `manifest.json` lists file sizes for download progress (including `webgpu/`).

Conversion script: `packages/tts/scripts/prepare-models.py` in the Seoga repository.

## License

The model weights are licensed under the **BigScience Open RAIL-M License** by Supertone Inc.
This repackaging is a derivative of the model and is distributed under the same license, including its use-based
restrictions (see `LICENSE`, Attachment A). Original model and code: [supertone-inc/supertonic](https://github.com/supertone-inc/supertonic)
(sample code MIT).