Voice Changer

Local, real-time AI voice changer for Ubuntu, controlled from the tray. Train a voice from your own audio files and use it as your microphone in Discord or any other app.

Requirements • Install • Usage • Soundboard • How it works

---

Voice conversion uses RVC v2 through Applio. Everything runs locally on your GPU, with no account or upload.

Requirements

• Ubuntu 24.04 (X11 session) with PipeWire
• NVIDIA GPU with CUDA (tested on an RTX 4060 8GB)
• uv

Install

./install.sh

Installs the apt packages it needs (ffmpeg libportaudio2 libxcb-cursor0 pipewire-bin), clones Applio at a pinned commit into vendor/, creates .venv, downloads the base models (~5 GB in total) and adds a "Voice Changer" launcher.

Usage

./run.sh

1. Tray → Open Voice Changer… → Voices tab → enter a name, choose audio files, then click Train. Use 10–30 min of clean speech with no music or noise. 300 epochs take about 30–60 min. Voice conversion pauses while training runs.
2. Tray → Voice → pick the new voice. Left-click the tray icon or press the hotkey (default Ctrl+Alt+V) to turn it on or off.
3. In Discord, set the input device to VoiceChanger Mic. Turn off Krisp noise suppression and echo cancellation, because they distort the converted voice.

When the voice changer is off, VoiceChanger Mic passes your normal voice through. Use the Monitor menu to hear the output in your headphones.

The tray menu also has:

• Pitch: shifts the converted voice from -12 to +12 semitones. Saved per voice.
• Latency: Low (100 ms), Balanced (200 ms, default) or Quality (350 ms). Larger chunks sound better but delay your voice more.
• Input device: which microphone to convert. Defaults to the system one.
• Start on login: adds the app to ~/.config/autostart.

Tray icon colors: green = on, gray = off, orange = loading or training, red = error.

Soundboard

Open the Soundboard tab in the window, or use the Soundboard menu in the tray. Click a sound to play it into VoiceChanger Mic. You also hear it on your default output, or through the Monitor when that is on. Playing a new sound stops the current one. Add sounds with Add sounds…, or drop audio files into data/sounds/. Right-click a sound to delete it. The volume slider applies to every sound.

How it works

 Microphone ──► RVC engine (Applio, GPU) ──► VoiceChanger Sink ──► VoiceChanger Mic ──► Discord
                                                  ▲     │
                       Soundboard (ffmpeg) ───────┘     └──► Monitor (headphones, optional)

The app creates a virtual sink/source pair with pw-loopback. The converted audio goes into VoiceChanger Sink, and PipeWire exposes it as VoiceChanger Mic, a regular input device any app can pick. When the voice changer is off or a voice is training, a second loopback sends the raw microphone straight to the sink, so the mic never goes silent. Soundboard clips are decoded by ffmpeg and played into the same sink with pw-cat.

Data

Everything lives in data/ (git-ignored): config.json, voices/<name>/ (model, index and pitch), datasets/<name>/ (copies of the uploaded audio) and sounds/ (soundboard files).

Tests

PYTHONPATH=src .venv/bin/python tests/test_hotkey.py

License

MIT. Applio, cloned into vendor/ during install, has its own license.