Applio App：桌面端语音转换工作室，支持模型训练与语音合成

A desktop-first, high-quality voice conversion studio built on Applio — convert voices, train models, and synthesize speech from one app.

🌐 Website • 📚 Documentation • ☎️ Discord

🛒 Plugins • 📦 Compiled • 🎮 Playground • 🔎 Google Colab • 🦅 Kaggle • 💻 Colab (No UI)

🎬 Applio Studio walkthrough showcase (Watch MP4 • WebM)

Introduction

Applio App is a voice conversion studio focused on simplicity, quality, and performance. Whether you're an artist, developer, or researcher, it gives you a single desktop app for high-quality voice transformations, model training, and speech synthesis — with a plugin system for extending it to your own projects.

Getting Started

Free Cloud Runtimes (Google Colab & Kaggle)

Run Applio online for free using cloud GPUs without installing anything locally:

• Google Colab: One-click setup with Nvidia T4 GPU, persistent Google Drive model synchronization, and free automatic Cloudflare tunneling (no account or token required).
• Kaggle Notebook: Free T4 x2 / P100 GPU with automatic Cloudflare tunneling (ensure Settings ➔ Internet: ON in Kaggle notebook settings).

Recommended: desktop installer

1. Download the installer for your OS from the releases page (Applio-<version>-Windows-Setup.exe, Applio-<version>-macOS.dmg, or Applio-<version>-Linux.AppImage / .deb) and run it.
2. Launch Applio. On first launch the setup screen checks every dependency — press Install / Repair and wait.
3. Every later launch re-runs the checks, so a broken environment is caught before you hit Convert.

Setup automatically downloads the default HiFi-GAN and RefineGAN training pretrains. With Use pretrained model enabled, training also downloads a missing pair for the selected vocoder and sample rate before starting. A failed download stops the job with an error; training from scratch requires explicitly disabling this option.

Store models and training logs on another drive

In Settings → Disk & Cache Storage → Logs & Models Folder, click Choose Folder to select a location on any drive, then Save Folder and close and reopen Applio. Downloaded models, training data, checkpoints, and TensorBoard logs use this folder. Click Use Default, then save, to restore the default location. In a browser, enter the folder path manually.

Applio automatically moves the existing folder on restart, including across drives, and updates training file lists to the new paths. It verifies the copied files before removing the original folder to free disk space. Existing destination files are preserved; conflicting files stop the move and leave the original folder in use.

Realtime latency

Realtime capture begins after the model is ready. Extra conversion context provides past audio history without imposing a startup silence period. The realtime status shows processing time and request round-trip time separately; microphone capture, playback, and audio-device buffering also contribute to the delay you hear.

Warning

macOS first launch: the app is not Apple-notarized yet, so Gatekeeper may claim Applio "is damaged". It isn't — clear the download quarantine and open it again:

xattr -cr /Applications/Applio.app

On Apple Silicon the app runs via PyTorch MPS with CPU fallback enabled automatically.

AMD GPU Support (Windows - ROCm)

Applio supports AMD GPUs on Windows out-of-the-box via native AMD ROCm PyTorch:

1. Launch Applio: Open Applio. On first launch, setup automatically detects your AMD GPU architecture.
2. Automatic Configuration: Press Install / Repair — Applio automatically downloads and installs native ROCm-accelerated PyTorch (torch[device-<target>]), torchaudio, and torchvision directly from AMD's official repository (https://stable.repo.amd.com/rocm/whl-next/).
3. Seamless Execution: Runs natively without translation layers or lengthy pre-compilation delays.

Supported AMD GPUs & Architectures

Developers (any OS)

Source layout:

frontend/web/      Web interface
frontend/desktop/  Electron shell and installers
backend/api/       App API and job orchestration
backend/applio/    Pinned shared RVC engine
uvr/              App-only audio separation
tools/            App-only audio utilities
assets/           Shared app resources

Prerequisites: Python 3.10–3.12, Node.js 22+, pnpm@11, ffmpeg.

git clone --recurse-submodules https://github.com/IAHispano/Applio-App.git
cd Applio-App
pnpm install
pnpm backend:init # also initializes the backend in existing checkouts
pnpm dev          # API :8000 + web :3000 with hot reload
# open http://localhost:3000/  (setup screen)

pnpm desktop:dev  # full Electron shell instead of just the browser UI

First run: open the app, press Install / Repair on the setup screen (creates .venv, installs torch + requirements, builds the UI).

RVC comes directly from the pinned Applio submodule. UVR stays in this app.

Audio separation installs its optional dependencies when first used, preserving the installed RVC runtime. Model weights stay in user storage rather than the app package.

Job history survives app restarts; interrupted work is marked so it can be started again. GPU jobs run one at a time, and realtime sessions hold new GPU work in the queue until stopped. Downloads recover from interrupted connections when the server supports resuming. Realtime shows processing time, estimated audio delay, queue depth, and dropped audio to help diagnose device and transport delays.

User settings

Preferences are stored separately from the app installation:

• Windows: %APPDATA%\Applio\config.json
• macOS: ~/Library/Application Support/Applio/config.json
• Linux: $XDG_CONFIG_HOME/Applio/config.json (defaults to ~/.config/Applio/config.json)

Desktop, web/API, and Python commands use the same settings. For a separate profile or a persistent cloud/container volume, set APPLIO_CONFIG_DIR to your preferred directory before launching Applio.

assets/config_template.json contains shipped defaults. Missing settings are filled from the current template. On first use, existing assets/config.json preferences are migrated if the per-user file does not exist; an existing per-user file is never replaced by migration. Only the template belongs in Git or release packages. Settings writes replace the file atomically, and invalid JSON is preserved rather than silently reset.

Scripts

CLI Mode

You can run Applio directly from your terminal using the applio command:

# Launch interactive terminal menu
applio

# Or execute specific commands directly:
applio infer --input-path input.wav --output-path out.wav --pth-path model.pth
applio batch-infer --input-folder inputs/ --output-folder outputs/ --pth-path model.pth
applio tts --tts-text "Hello world" --output-tts-path tts.wav
applio uvr --input-path song.mp3 --output-dir stems/
applio train --model-name my_voice
applio download --model-link <url>
applio web
applio --help

Terms of Use and Commercial Usage

Using Applio responsibly is essential.

• Users must respect copyrights, intellectual property, and privacy rights.
• Applio is intended for lawful and ethical purposes, including personal, academic, and investigative projects.
• Commercial usage is permitted, provided users adhere to legal and ethical guidelines, secure appropriate rights and permissions, and comply with the MIT license.

The source code and model weights in this repository are licensed under the permissive MIT license, allowing modification, redistribution, and commercial use.

However, if you choose to use this official version of Applio (as provided in this repository, without significant modification), you must also comply with our Terms of Use. These terms apply to our integrations, configurations, and default project behavior, and are intended to ensure responsible and ethical use without limiting their use in any way.

For commercial use, we recommend contacting us at support@applio.org to ensure your usage aligns with ethical standards. All audio generated with Applio must comply with applicable copyright laws. If you find Applio helpful, consider supporting its development through a donation.

By using the official version of Applio, you accept full responsibility for complying with both the MIT license and our Terms of Use. Applio and its contributors are not liable for misuse. For full legal details, see the Terms of Use.

References

Applio is made possible thanks to these projects and their references:

• gradio-screen-recorder by gstaff
• rvc-cli by blaisewf

Contributors