mlx-tts-studio

KO | EN

gitlite — search

#python #ai #golang #android #deep-learning #docker #macos #java #windows #vue #pytorch #security #ios #mcp

(★ 13)

Native macOS text-to-speech app powered by Qwen3-TTS and Apple Silicon (MLX). Voice cloning, voice design, and custom voices — all running locally.

mlx-tts-studio Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# MLX TTS Studio

A native macOS desktop app for high-quality text-to-speech using [Qwen3-TTS](https://huggingface.co/Qwen/Qwen3-TTS) models, accelerated by Apple Silicon via [MLX](https://github.com/ml-explore/mlx).

Built with Tauri v2, React 19, and a Python inference engine running locally on your Mac. No cloud, no API keys, no data leaves your machine.

![Generate Page](https://raw.githubusercontent.com/louiscoetzee/mlx-tts-studio/main/screenshots/app-generate-page.png)

## Features

**Three TTS modes:**

- **Custom Voice** -- Choose from preset speakers across English, Chinese, Japanese, and Korean. Add emotion/style instructions like "excited and happy" or "calm narrator".
- **Voice Design** -- Describe the voice you want in natural language (e.g. "warm British male narrator in his 40s") and the model creates it.
- **Voice Clone** -- Clone any voice from a short audio sample. Upload a file, record from your microphone, or use a saved voice profile.

**Model management:**

- **Pro (1.7B)** and **Lite (0.6B)** model tiers -- higher quality vs. faster generation
- One-click download from HuggingFace with progress tracking
- 6 specialized models (Custom Voice / Voice Design / Voice Clone for each tier)

**Audio:**

- Waveform visualization with wavesurfer.js
- Playback controls, volume, speed adjustment (0.5x--2.0x)
- Save generated audio as WAV
- Generation progress with cancel support

**Voice library:**

- Enroll voices from recordings or uploaded files
- Manage saved voice profiles for reuse in Voice Clone mode

**Everything runs locally:**

- Guided first-run setup installs Python environment automatically
- Models download from HuggingFace and run on-device via MLX
- All data stored in `~/.qwen-tts/`

<details>
<summary>More screenshots</summary>

| Voice Clone | Models |
|:-----------:|:------:|
| ![Voice Clone](https://raw.githubusercontent.com/louiscoetzee/mlx-tts-studio/main/screenshots/app-voice-clone-tab.png) | ![Models](https://raw.githubusercontent.com/louiscoetzee/mlx-tts-studio/main/screenshots/app-first-launch.png) |

| Voices | Settings |
|:------:|:--------:|
| ![Voices](https://raw.githubusercontent.com/louiscoetzee/mlx-tts-studio/main/screenshots/app-voices-page.png) | ![Settings](https://raw.githubusercontent.com/louiscoetzee/mlx-tts-studio/main/screenshots/app-settings-page.png) |

</details>

## Requirements

- **macOS 14.0+** (Sonoma or later)
- **Apple Silicon** (M1/M2/M3/M4) -- required for MLX acceleration
- **8GB+ RAM** (16GB recommended for Pro models)
- [Homebrew](https://brew.sh)

## Getting Started

### 1. Install system dependencies

```bash
# Rust toolchain
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# Node.js (v18+)
brew install node

# ffmpeg (for audio format conversion)
brew install ffmpeg

# Python 3.11+ and uv (fast package manager)
brew install python@3.11 uv
```

### 2. Clone and install

```bash
git clone https://github.com/YOUR_USERNAME/mlx-tts-studio.git
cd mlx-tts-studio
npm install
```

### 3. Run in development mode

```bash
npm run tauri dev
```

On first launch, the app will guide you through setting up the Python environment and downloading a model.

### 4. Build for production

```bash
npm run tauri build
```

The `.dmg` installer will be in `src-tauri/target/release/bundle/dmg/`.

## Usage

1. **Start the engine** -- Click "Start" in the sidebar footer to launch the Python TTS engine.
2. **Download a model** -- Go to Models and download at least one model for the mode you want to use.
3. **Generate speech** -- Type or paste text on the Generate page, configure voice settings, and click Generate.
4. **Save output** -- Use the "Save As" button to export the generated WAV file.

### Voice Clone workflow

1. Go to the **Voice Clone** tab
2. Provide reference audio (upload a file, record from mic, or select a saved voice)
3. Type the **exact transcript** of what's said in the reference audio
4. Enter the text you want spoken in the cloned voice
5. Click Generate

For best results, use a clean 5--15 second recording with minimal background noise.

## Models

| Tier | Parameters | VRAM | Quality | Speed |
|------|-----------|------|---------|-------|
| **Pro** | 1.7B (8-bit) | ~3.3 GB | Higher | Slower |
| **Lite** | 0.6B (8-bit) | ~1.3 GB | Good | Faster |

Each tier has three specialized models (Custom Voice, Voice Design, Voice Clone). The app downloads and manages them for you -- just pick Pro or Lite on the Generate page.

Models are sourced from [mlx-community](https://huggingface.co/mlx-community) on HuggingFace, quantized to 8-bit for efficient Apple Silicon inference.

## Architecture

```
┌─────────────────────────────────────────────────┐
│                   Tauri v2 App                  │
│  ┌──────────────┐          ┌─────────────────┐  │
│  │   React 19   │  IPC     │   Rust Backend  │  │
│  │   Frontend   │◄────────►│   (Tokio)       │  │
│  │              │          │                 │  │
│  │  - Generate  │          │  - Model mgmt   │  │
│  │  - Models    │          │  - Downloads     │  │
│  │  - Voices    │          │  - Audio I/O     │  │
│  │  - Settings  │          │  - Voice library │  │
│  └──────────────┘          └────────┬────────┘  │
│                                     │ HTTP      │
│                            ┌────────▼────────┐  │
│                            │  Python Engine  │  │
│                            │  (Flask + MLX)  │  │
│                            │                 │  │
│                            │  - TTS Inference│  │
│                            │  - Model loading│  │
│                            │  - SSE streaming│  │
│                            └─────────────────┘  │
└─────────────────────────────────────────────────┘
```

- **React frontend** handles UI, state management (Zustand), and user interactions
- **Rust backend** manages the filesystem, model downloads, audio processing, and spawns the Python engine as a subprocess
- **Python engine** runs the actual TTS inference using mlx-audio on Apple Silicon, communicating via a local HTTP/SSE API

## Tech Stack

| Layer | Technology |
|-------|-----------|
| Desktop framework | [Tauri v2](https://v2.tauri.app) |
| Frontend | React 19, TypeScript, Tailwind CSS 4 |
| UI components | shadcn/ui (Radix UI) |
| State management | Zustand 5 |
| Audio visualization | wavesurfer.js 7 |
| Backend | Rust, Tokio |
| TTS inference | [mlx-audio](https://github.com/Blaizzy/mlx-audio), MLX |
| Models | [Qwen3-TTS](https://huggingface.co/Qwen/Qwen3-TTS) (8-bit quantized) |

## Data Storage

All app data lives in `~/.qwen-tts/`:

```
~/.qwen-tts/
├── models/       # Downloaded model weights
├── voices/       # Enrolled voice profiles
├── outputs/      # Generated audio files
├── python-env/   # Managed Python virtualenv
├── config.json   # App settings
└── history.json  # Generation history
```

## Development

### Project structure

```
qwen3-tts-studio/
├── src/                    # React frontend
│   ├── pages/              # Generate, Models, Voices, Settings, Setup
│   ├── components/         # UI components
│   ├── lib/                # Tauri API wrappers, types, utilities
│   └── store/              # Zustand state management
├── src-tauri/              # Rust backend
│   ├── src/                # Commands, models, downloads, audio, etc.
│   └── python-engine/      # Flask TTS server + requirements.txt
├── package.json
├── vite.config.ts
└── tsconfig.json
```

### Useful commands

```bash
# Frontend only (no Tauri)
npm run dev

# Full app with hot reload
npm run tauri dev

# Type check
npx tsc --noEmit

# Production build
npm run tauri build
```

## Acknowledgments

- [qwen3-tts-apple-silicon](https://github.com/kapi2800/qwen3-tts-apple-silicon) by kapi2800 -- the CLI tool that served as the reference and inspiration for this project. The Python inference engine is based on their work.
- [Qwen3-TTS](https://huggingface.co/Qwen/Qwen3-TTS) by Alibaba DAMO Academy
- [mlx-audio](https://github.com/Blaizzy/mlx-audio) by Prince Canuma
- [MLX](https://github.com/ml-explore/mlx) by Apple
- [Tauri](https://tauri.app) by the Tauri contributors

## License

MIT

mlx-tts-studio

mlx-tts-studio Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

mlx-tts-studio

mlx-tts-studio Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README