Skip to content

About

Realtime voice changer for Linux: natural pitch shifting, effects, presets and local AI voices, as a PipeWire microphone and a CLAP plugin

Topics

Resources

Contributing

Stars

3 stars

Watchers

0 watching

Forks

Repository files navigation

Voice Changer

A realtime voice changer for Linux that shows up as a microphone called Voice Changer Mic. Natural pitch shifting, effects, presets and local AI voices. Discord, OBS and your softphone just pick the mic.

Latest release CI Downloads MIT license

Voice Changer home page with preset cards, quick-tweak knobs and the live meters

What is it?

Voice Changer sits between your microphone and everything else. It cleans the signal up, changes the voice, and publishes the result as a new PipeWire microphone. Any app that can pick a microphone can use it. There is nothing to configure in PulseAudio or JACK.

  • Sounds like a person. The default Natural pitch engine is pitch-synchronous (the technique hardware voice transformers use), so shifted voices stay human instead of turning into a vocoder. Pitch and formant ("character") move independently.
  • Ready-made voices. Man, Woman, Baritone, Bass, Deep, Demon, Monster, Chipmunk, Robot, Radio, Cave, Announcer. Star your favourites, save your own.
  • AI voices, locally. Load any RVC v2 voice model and become that voice. Runs on the CPU or an AMD GPU, nothing leaves your machine. Retrieval indices are supported for a closer match.
  • Clean input. Neural noise suppression (RNNoise) and a voice-only gate run before everything else, so fans, keyboards and breaths do not get turned into voice.
  • One big switch. ON/OFF in the window, in the tray, on a global shortcut, or from the terminal with voice-changer toggle.
  • Hear myself. One click plays the processed voice to your headphones.
  • Also a plugin. The same engine and interface load as a CLAP plugin in your DAW.

Voices page: importer and installed voice list Effects page: pitch engine and effect panels with knobs

Mixer page: gains, auto level, noise suppression, gates and limiter Settings page: microphone, window behaviour, hotkey, theme and log

Install

Fedora, openSUSE, RHEL family Download the .rpm from the latest release, then sudo dnf install ./voice-changer-*.rpm
Debian, Ubuntu, Mint, Zorin, Pop!_OS Download the .deb, then sudo apt install ./voice-changer_*.deb
Arch, Manjaro, EndeavourOS git clone this repository, cd packaging/aur, makepkg -si
Any Linux (also Steam Deck desktop mode) Download VoiceChanger.flatpak, then flatpak install --user ./VoiceChanger.flatpak; or unpack the .tar.gz and run ./voice-changer (the CLAP plugin is in the same folder)
From source See Building. scripts/install.sh builds and installs everything for your user.

Every asset comes for x86_64 and aarch64 (Raspberry Pi 5, Asahi Linux, ARM laptops), with a SHA256SUMS file next to them.

AI voices run on the processor everywhere. AMD GPUs are used out of the box with the onnxruntime-rocm package; NVIDIA GPUs after voice-changer get-runtime --cuda (see Compute). voice-changer doctor tells you what your machine has and what is missing.

The packages install the app, the CLAP plugin (/usr/lib64/clap/ or /usr/lib/clap/), a launcher and the icon. AI voice models are downloaded separately (see AI voices).

Requirements

Sound server PipeWire with WirePlumber, as the audio server (not only for screen sharing). Tested with PipeWire 1.0 and 1.6.
System glibc 2.35 or newer: Ubuntu 22.04+, Debian 12+, Fedora 36+, Linux Mint 21+, Zorin 17+, openSUSE Leap 15.6+ / Tumbleweed, Arch, NixOS
Desktop Wayland or X11. Any GPU; virtual machines fall back to software rendering
Optional A tray (KDE, GNOME with the AppIndicator extension), the GlobalShortcuts portal, pipewire-pulse so PulseAudio apps (Discord, browsers, OBS) see the microphone

Ubuntu 22.04 and Debian 12 ship PulseAudio as the default sound server; install pipewire-audio (Debian: pipewire-audio or pipewire-pulse + wireplumber), log out and in, and you are set. voice-changer doctor tells you exactly what is missing and what to install.

Using it

  1. Start Voice Changer from your launcher (or run voice-changer).
  2. Pick a preset on Home, or turn the quick-tweak knobs.
  3. In Discord, OBS, your softphone, choose Voice Changer Mic as the input device.
  4. Tick Hear myself to check how you sound. Use headphones.
Where What
Home Preset cards with search and favourites, quick tweaks (Pitch, Character, Grit, Space, Echo), the AI voice card
Voices Import AI voices from a link, list and delete installed ones, download the base models
Effects Pitch engine, Voice, Drive, Robot, Echo, Reverb, Tone
Mixer Input and output gain, auto level, limiter, noise suppression, voice-only gate, noise gate
Settings Microphone, hear-myself, audio status, close to tray, start minimized, autostart, hotkey, AI threads and engine, theme, log

Knobs: drag up or down, roll the mouse wheel, double-click or right-click to go back to the default. Sliders: drag or click anywhere on the rail, wheel, and the same double-click or right-click reset. Hover any control for a moment and it explains itself (the top bar is left alone; it speaks for itself).

Where the voice goes. Any app can pick Voice Changer Mic in its own settings. The Mixer page's Send the voice to section does it for you: Use as the microphone for everything makes it the system default while the app runs (your previous default comes back when you turn it off or quit), and the list below sends it to individual recording applications, Discord, OBS, a browser tab, without touching anything else.

Shortcut Action
Ctrl+M or Ctrl+Space Toggle on/off (window focused)
Ctrl+H Hear myself
Ctrl+1 … Ctrl+5 Home, Voices, Effects, Mixer, Settings
Ctrl+, Settings
Ctrl+F Search presets
Ctrl+S Save the current settings as a preset
Esc Cancel a save or delete prompt
Ctrl+Q Quit
Ctrl+Shift+M Global toggle, from any app (desktop portal; change it under Settings → On/off hotkey)

Closing the window quits and saves your state. Turn on Close to tray in Settings if you would rather have it keep running in the tray.

From the terminal

Everything the window does is also a command against the running instance:

voice-changer                      # window + tray + virtual mic
voice-changer toggle               # flip on/off (bind this to a key if the portal shortcut is unavailable)
voice-changer on | off | status
voice-changer preset Robot         # apply a preset; `presets` lists them
voice-changer mics                 # list microphones; `mic <node.name>` switches, `mic default` resets
voice-changer monitor on           # hear yourself; `monitor off`
voice-changer apps                 # who is recording and whether they get the voice
voice-changer route Discord on     # send the voice to an application; `route Discord off`
voice-changer default on           # be the default microphone for everything; `default off`
voice-changer ai "Nice Voice"      # switch on an AI voice; `ai off`; `ai-voices` lists them
voice-changer voices               # installed voice models
voice-changer voices add https://huggingface.co/user/voice-model "Nice Voice"
voice-changer voices remove "Nice Voice"
voice-changer show | quit
voice-changer doctor               # check this computer: PipeWire, session manager, mics, GPU, models, tray
voice-changer --self-test          # run the engine through every preset (no audio devices needed)
voice-changer --quantum 512        # bigger audio block if you hear dropouts
voice-changer --ai-threads 4       # AI inference threads (default: auto, from your CPU)
voice-changer --help

AI voices

Any RVC v2 model works, which is the format almost every community voice comes in.

  1. Base models. The Voices page offers a one-click download of ContentVec and RMVPE (MIT licensed, about 740 MB), or run scripts/get-models.sh.
  2. Add a voice. Paste a voice-models.com voice page, a Hugging Face model page or a direct .pth / .onnx / .zip link into Voices → Add a voice. .pth files are converted to ONNX automatically (the first conversion sets up a small Python environment, about 300 MB). A matching .index file is picked up when the repository or zip has one.
  3. Use it. Press Use on the voice, or pick it on the Home page's AI card.

Voices with more than one speaker. Some models are trained on two takes, say a male and a female version, selected by RVC's speaker id. RVC files do not record how many speakers a model has, so on the Voices page set Speakers to 2 (or more) for that voice; a Speaker picker then appears on the Home page, switching is instant, and Starts as pins the speaker the voice always begins with. You can name the speakers ("Male", "Female") right there. From the terminal: voice-changer ai "Nice Voice" 2.

Where to find voices: Hugging Face (search "RVC"), voice-models.com, the AI Hub community. Check each model's licence and the voice owner's consent before using a voice publicly. The app ships no voices.

Tips for the most natural result:

  • AI pitch matters most. Set it so the converted voice lands in the model's natural range: +12 for a typical male→female model, 0 or −12 the other way. Leave the DSP Pitch at 0.
  • Breathiness scales the model's noise excitation. Lower for a cleaner voice, higher for more texture.
  • Index strength appears when the voice has a retrieval index. Around 75% keeps the model's timbre; lower if words become unclear.
  • AI only (on by default) bypasses the DSP voice and effects while AI runs. Turn it off to layer Robot, reverb or extra pitch on top.
  • Prefer models trained for 300+ epochs on an hour or more of clean speech, at 40k or 48k.
  • Protect consonants (with an index) keeps breaths and consonants closer to what you said instead of pulling them toward the index. Raise it if s, t and sh sound smeared.
  • Follow my loudness makes the converted voice as loud, moment by moment, as what you put in, so quiet breaths and mouth noises stay quiet and the noise gate works on the AI voice too.
  • Block coughs and burps silences coughs, burps, sneezes, throat clearing and hiccups instead of converting them (laughter is left to the voice). It uses a small sound classifier, YAMNet (15 MB, Apache-2.0), fetched with the base models.
  • Settings belong to the voice. Pitch, index, breathiness, protection and loudness are remembered per voice and come back when you select it.
  • Test bench (Voices page): record a phrase once and replay it through any voice while you move the sliders. A replayed clip is heard through Hear myself only; the virtual microphone is silent while it plays.
  • Check on a voice converts a test phrase and reports clipping, silence, poor pitch tracking or a missing index. New imports are checked automatically.
  • AI speed (Auto) starts at the fastest setting and backs off to longer blocks if your machine falls behind. Expect about 0.5 s of delay on a GPU and 0.65 s on the processor; it needs a block of speech before it can convert it.

Compute: CPU or GPU

Compute on the AI card picks Auto, CPU or a specific GPU; the models reload when you change it. On the CPU expect about 60–80% of eight cores on a Ryzen 5950X for a 48 kHz model. On an AMD GPU the content and pitch models run 4–5× faster and the voice sees a full second of context per block, which keeps the timbre steadier.

The app loads ONNX Runtime at start and uses whichever it finds first:

Runtime Where it comes from Result
ROCm build Fedora: sudo dnf install onnxruntime-rocm (or ORT_DYLIB_PATH) AMD GPUs appear under Compute
CUDA build voice-changer get-runtime --cuda (needs NVIDIA's driver, CUDA 12 runtime and cuDNN 9) NVIDIA GPUs appear under Compute
CPU build bundled with the packages, voice-changer get-runtime, or the distribution's onnxruntime CPU only

Settings → AI shows which library is loaded. The first load of a voice on an AMD GPU compiles kernels, which takes tens of seconds and is cached in ~/.cache/miopen.

Threads. The AI picks its thread count from your processor: physical cores minus one (so audio and the window always have a core), only the performance cores on hybrid chips, at most 12, and just a few when the models run on a GPU. Settings → AI shows what was detected and lets you choose a number up to the threads you have; --ai-threads N does the same from the terminal. Lower it if other programs stutter while the AI runs.

NVIDIA. Nothing is needed for the app, the effects or AI voices on the processor. To run the AI on the graphics card:

  1. Install the driver and the CUDA 12 runtime with cuDNN 9. Fedora: RPM Fusion's xorg-x11-drv-nvidia-cuda plus cuda-toolkit-12 and cudnn9 from NVIDIA's repository. Ubuntu: nvidia-driver-5xx, cuda-toolkit-12-x, cudnn9-cuda-12. Arch: cuda, cudnn.
  2. voice-changer get-runtime --cuda downloads Microsoft's CUDA build of ONNX Runtime into ~/.local/lib/voice-changer/cuda.
  3. voice-changer doctor says whether everything was found; then start the app and pick the GPU under Compute (Auto chooses it).

If a piece is missing the AI runs on the processor and Settings → AI says which piece. The same binary serves AMD, NVIDIA and CPU-only machines.

Plugin

cargo xtask bundle vc-plugin --release writes target/bundled/vc-plugin.clap; the packages install it system-wide and scripts/install.sh puts it in ~/.clap/. The editor is the same interface as the app, minus the microphone and window settings. VST3 can be built with --features vst3 (mind Steinberg's licensing terms).

Latency

Path Added delay at 48 kHz, 256-frame block
Natural engine ≈ 45 ms engine + one block (5 ms)
Fast / Balanced / Smooth ≈ 20–60 ms depending on mode
Noise suppression + 10 ms
AI voice + 0.5–1 s (needs a block of speech to convert)

The current figure is shown in the sidebar. --quantum 128 shaves 2.7 ms if your interface copes; USB headsets usually do not.

Troubleshooting

Start with voice-changer doctor: it checks the sound server, the session manager, virtual-device support, microphones, GPU and AI runtime, models, tray and shortcut portal, and prints the fix for anything missing.

  • "Audio is not running" in the window. PipeWire is not the sound server on this machine, or is not running. Ubuntu 22.04 / Debian 12: sudo apt install pipewire-audio (or pipewire-pulse wireplumber), log out and in. Elsewhere: systemctl --user start pipewire wireplumber pipewire-pulse. Press Try again.
  • The app looks right but there is no "Voice Changer Mic" in Discord or the browser. Those apps talk PulseAudio; make sure pipewire-pulse is installed and pactl info says Server Name: PulseAudio (on PipeWire …).
  • Blank, frozen or very slow window (VMs, some NVIDIA driver and Wayland combinations). Run with WGPU_BACKEND=gl voice-changer (OpenGL instead of Vulkan) or ICED_BACKEND=tiny-skia voice-changer (software rendering). The audio side is not affected by either.
  • The app closes when an AI voice is picked (0.3.1 and older, on machines without an AMD GPU). Fixed in 0.4.0; update.
  • No window and no tray icon. GNOME needs the AppIndicator extension for a tray; the app opens its window instead when "Start minimized" finds no tray. voice-changer show always brings it up.
  • Pops or crackle. Check the sidebar for "dropouts". Start with a bigger block (voice-changer --quantum 512), then run pw-top while speaking: an ERR count rising on your microphone or on voice_changer.* nodes confirms PipeWire xruns.
  • Words get cut. Turn off Voice only in Mixer, or raise Voice Gate Depth (e.g. −20 dB) so more passes between words.
  • No "Voice Changer Mic" in an app. The node exists only while the app runs. Flatpak apps need PipeWire access (--filesystem=xdg-run/pipewire-0).
  • The app runs, the meters move, but nobody hears the changed voice. The other app is still using your real microphone. Either pick Voice Changer Mic in that app's own input settings, tick the app under Mixer → Send the voice to, or turn on Use as the microphone for everything and leave the app on "Default". wpctl status shows which source each recording stream is linked to.
  • Changed voice is very quiet. Older session managers remembered a low volume for the virtual mic. Since 0.2.2 the app holds it at 100% itself; on older versions run wpctl set-volume <id of Voice Changer Mic> 1.0.
  • Importing a .pth fails with "No module named …". Update to 0.2.1 or later; the converter environment repairs itself on the next import.
  • Global shortcut does nothing. Your desktop may lack the GlobalShortcuts portal. Bind voice-changer toggle to a key in System Settings → Shortcuts → Custom instead.
  • AI "error" or "loading" forever. Settings → Log shows why. Common causes: base models missing (Voices page), a v1 model (256-dim) instead of v2, or a GPU runtime that is not installed (sudo dnf install onnxruntime-rocm).
  • Test without talking. voice-changer --input-wav speech.wav --monitor loops a file as the microphone.

Where your stuff is kept

Settings and last state ~/.config/voice-changer/config.ron, state.ron
Launcher and icons (written on first start unless a package installed them) ~/.local/share/applications/, ~/.local/share/icons/hicolor/
CPU ONNX Runtime (when fetched by the install script) ~/.local/lib/voice-changer/
Your presets and favourites ~/.config/voice-changer/presets/*.ron, favorites.txt
Base models and voices ~/.local/share/voice-changer/models/, models/voices/
Converter environment ~/.local/share/voice-changer/tools/

Open data folder on the Settings page takes you there.

Building

# Fedora
sudo dnf install pipewire-devel clang gcc-c++ libxkbcommon-devel wayland-devel
# Debian / Ubuntu
sudo apt install libpipewire-0.3-dev libclang-dev pkg-config libxkbcommon-dev libwayland-dev build-essential

git clone https://github.com/ZeroDay-Labz/voice-changer && cd voice-changer
scripts/install.sh              # app + CLAP plugin + launcher into ~/.local (GPU-capable)
scripts/install.sh --cpu-only   # without the dynamic runtime loader
cargo run                       # or just run it from the tree

Rust 1.88 or newer. cargo test --workspace runs the DSP and index tests; cargo clippy --workspace --all-targets must stay clean. See CONTRIBUTING.md.

Crate Purpose
vc-dsp Realtime-safe engine (no allocation or locks in process)
vc-core Parameters, presets, pipeline, AI inference and voice library
vc-gui iced interface shared by the app and the plugin
vc-app Standalone: PipeWire virtual mic, window, tray, hotkey, D-Bus, CLI
vc-plugin CLAP (and optional VST3) plugin via nice-plug

Notes

  • Mono, 48 kHz internally; multichannel plugin inputs are folded to mono and fanned out.
  • The D-Bus interface is org.voicechanger.Control on the session bus; the CLI is a thin client for it.
  • ONNX Runtime upstream dropped its ROCm build; the app loads the distribution's onnxruntime-rocm library at runtime instead, and Microsoft's CUDA build for NVIDIA.
  • Screenshots in this README are rendered by the app itself: voice-changer --screenshot docs/screenshots/home.png --page home.

License

MIT. See LICENSE. Voice models you download have their own licences.

About

Realtime voice changer for Linux: natural pitch shifting, effects, presets and local AI voices, as a PipeWire microphone and a CLAP plugin

Topics

Resources

Contributing

Stars

3 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages