Supertonic 3 — TTS App

On-device neural text-to-speech for Windows, powered by Supertone's Supertonic 3 engine.

<p align="center"><strong>English</strong> | <a href="README_ZH.md">中文</a></p>
</p>

What It Does

Supertonic 3 TTS App is a fully offline text-to-speech application that runs neural TTS inference entirely on your device. No cloud, no API keys, no data leaves your machine.

It wraps the Supertone Supertonic 3 ONNX models inside a Tauri 2 desktop app with a premium two-column UI built with React and TypeScript. Audio generation uses ONNX Runtime Web with GPU acceleration via WebGPU and a multi-threaded WASM CPU fallback.

Key Features

Architecture

┌─────────────────────────────────────────────────┐
│                   Tauri 2 Shell                 │
│  ┌──────────────────────┐  ┌─────────────────┐  │
│  │   React + TypeScript │  │   Rust Backend   │  │
│  │                      │  │                  │  │
│  │  • UI & State        │  │  • File I/O      │  │
│  │  • ONNX Runtime Web  │  │  • PDF extraction│  │
│  │  • WebGPU / WASM     │  │  • DOCX parsing  │  │
│  │  • Audio playback    │  │  • Save dialogs  │  │
│  │  • MP3 encoding      │  │  • Clipboard     │  │
│  └──────────────────────┘  └─────────────────┘  │
└─────────────────────────────────────────────────┘

Frontend handles all ML inference, audio processing, and encoding in-browser via ONNX Runtime Web + lamejs-fixed.

Rust backend handles filesystem access (read files, save audio, extract text from PDF/DOCX), clipboard, native dialogs, WebView headers for SharedArrayBuffer, and logical CPU thread detection. Zero sidecars, zero Python, zero FFmpeg.

Prerequisites

• Node.js 18+ and npm
• Rust 1.70+ (rustup)
• Tauri CLI v2 (npm install @tauri-apps/cli)
• UPX (optional, for compression) — upx.github.io
• Models — download from HuggingFace Supertone/supertonic-3 (see below)

Getting Started

Want the prebuilt app? Download the latest release — portable EXE or NSIS installer, no build tools needed.

1. Clone

git clone https://github.com/Saganaki22/Supertonic_3_TTS_App.git
cd Supertonic_3_TTS_App

2. Install dependencies

npm install

3. Download models

The app fetches models at runtime from HuggingFace on first load and caches them in the browser. No manual download required — but you need an internet connection the first time you load a voice.

The following 6 files are fetched and SHA-256 verified:

Models are sourced from HuggingFace Supertone/supertonic-3.

4. Development

npm run tauri dev

Opens the app in development mode with hot-reload.

5. Build

npm run tauri build

Produces a portable NSIS installer at:

src-tauri/target/release/bundle/nsis/Supertonic 3 TTS App_0.1.4_x64-setup.exe

And the raw executable at:

src-tauri/target/release/supertonic-3-tts-app.exe

6. Compress (optional)

Using UPX:

upx --best --lzma src-tauri/target/release/supertonic-3-tts-app.exe

This typically reduces the exe from ~12 MB to ~7.2 MB.

Project Structure

├── index.html                  # Vite entry point
├── package.json                # npm config
├── tsconfig.json               # TypeScript config
├── vite.config.ts              # Vite config
├── public/
│   ├── assets/logo.png         # App logo
│   └── voice-samples/          # Voice sample WAVs (F1–F5, M1–M5)
├── scripts/
│   └── copy-ort-assets.js      # Copies ONNX Runtime Web assets for dev/prod
├── src/
│   ├── main.tsx                # Entry point, global event handlers
│   ├── App.tsx                 # Root component, layout, state
│   ├── App.css                 # All styles (themes, components)
│   ├── vite-env.d.ts           # Vite type declarations
│   ├── lamejs.d.ts             # Type declarations for lamejs-fixed
│   ├── components/
│   │   ├── Header.tsx          # Logo, brand, settings, status badge
│   │   ├── Footer.tsx          # Version, GitHub link
│   │   ├── SettingsPanel.tsx   # Theme, accent, provider, CPU usage, volume, UI scale
│   │   ├── StatusBadge.tsx     # Model status + GPU/CPU chevron dropdown
│   │   ├── TextInput.tsx       # Textarea with char count, paste, browse
│   │   ├── VoiceGrid.tsx       # 10-voice selector grid
│   │   ├── VoiceControls.tsx   # Steps & speed sliders
│   │   ├── VoiceClone.tsx      # Custom voice .json upload
│   │   ├── OutputPanel.tsx     # Waveform, player, timestamped WAV/MP3 export
│   │   ├── AudioHistory.tsx    # Up to 6 previous generations
│   │   ├── PanelResize.tsx     # Resizable panel divider
│   │   ├── CachePanel.tsx      # Model cache info
│   │   ├── ProgressBar.tsx     # Generation progress bar
│   │   └── Toast.tsx           # Toast notifications
│   ├── hooks/
│   │   ├── useTTS.ts           # ONNX engine state management
│   │   ├── useAudioPlayer.ts   # Web Audio playback + seeking
│   │   └── useAudioHistory.ts  # Six-item audio history
│   └── lib/
│       ├── tts.ts              # ONNX inference engine, explicit ORT assets, SHA-256
│       ├── helpers.ts          # chunkText, normalizeAudio, wavToMp3
│       └── tauri.ts            # Typed Tauri invoke wrappers
└── src-tauri/
    ├── Cargo.toml              # Rust dependencies
    ├── tauri.conf.json         # Tauri window, bundle, CSP config
    ├── build.rs                # Tauri build script
    ├── capabilities/
    │   └── default.json        # Tauri permission capabilities
    ├── icons/                  # App icons (ico, png)
    └── src/
        ├── main.rs             # Rust entry point
        ├── lib.rs              # Plugin registration, WebView headers, window setup
        └── commands/
            ├── mod.rs          # Command module exports
            ├── fs.rs           # read_text_file, save_wav/mp3_file
            ├── pdf.rs          # extract_pdf_text (20 MB limit)
            ├── docx.rs         # extract_docx_text (20 MB limit)
            └── system.rs       # logical_cpu_count for CPU thread auto-detection

Configuration

All settings are accessible from the gear icon in the header:

Settings persist in localStorage. Window position and size persist via tauri-plugin-window-state.

Custom Voice Styles

Train your own voice embeddings using the companion trainer:

Saganaki22/supertonic_embeddings_trainer

Export a .json file containing style_ttl and style_dp tensors, then load it via the "Custom voice" section in the app. The file is validated on load (JSON schema, tensor dimensions, data length).

Voice Samples

Built-in voices are English neural TTS voices. Sample audio files are in public/voice-samples/:

• F1–F5: Female voices (English)
• M1–M5: Male voices (English)

Tech Stack

License

This project (application code) is licensed under the MIT License.

The accompanying TTS model is released under the OpenRAIL-M License. See the LICENSE file in the model repository for details.

The TTS engine is developed by Supertone. Models are fetched from HuggingFace Supertone/supertonic-3.