Voicebox

Real-time voice pipeline for macOS:

Microphone → DeepFilterNet (denoise) → RVC (voice change) → BlackHole 2ch

Any app that lets you pick a microphone — Discord, Zoom, OBS, Google Meet, QuickTime — will receive the processed voice when its mic input is set to BlackHole 2ch.

Built with PySide6 on top of deepfilternet, rvc-python, fairseq, torch (MPS on Apple Silicon, CPU fallback for unsupported ops). Apple Silicon native; Intel macOS works on CPU.

---

What it does

• Low-latency mic capture at 48 kHz mono via CoreAudio (sounddevice).
• Real-time denoise with DeepFilterNet3.
• Real-time voice conversion with RVC v2:
  • HuBERT or ContentVec content encoder (ContentVec recommended for non-English speech).
  • rmvpe F0 extraction with configurable pitch shift, filter radius, protect, and index rate.
  • Streaming sliding-window inference with equal-power crossfade.
• Routing to BlackHole 2ch (virtual audio loopback), so any app can pick the processed voice up as a mic.
• Native macOS UX: frameless window with custom title bar, menu-bar tray icon (live indicator), close-to-tray, dock icon, mic permission prompt, Audio MIDI Setup walkthrough.

Installation (end users)

Download the .dmg from the Releases page (or build it yourself, see below):

1. Open Voicebox.dmg, drag Voicebox.app to /Applications.
2. First launch: Right-click Voicebox in /Applications → Open → Open (the bundle is unsigned; macOS Gatekeeper asks once).
3. Open the app, hit Setup, click Download next to Base models (~350 MB).
4. (Optional, recommended for non-English voices) Click Download ContentVec (multilingual) under Voice models → Encoder.
5. Install BlackHole 2ch if you haven't already: Existential Audio · BlackHole or brew install blackhole-2ch.
6. In Audio MIDI Setup, create a Multi-Output Device (BlackHole + your speakers) so you can hear yourself while routing audio to BlackHole. Voicebox shows a step-by-step guide.
7. Drop a .pth + .index voice model into the Voice models drop zone, or paste a Hugging Face / direct URL.
8. Open Voice Pipeline, pick input/output devices and RVC model, hit Start pipeline.
9. In your target app (Discord/Zoom/OBS), set the microphone input to BlackHole 2ch. Keep the output (speakers/headphones) on your normal device.

Find more voice models on weights.gg or Hugging Face.

Pipeline knobs (Voice tuning)

• Pitch shift (−24..+24 semitones)
• Protect (0.00..0.50) — RVC's voiceless-consonant protection. Lower = more original consonants preserved.
• Filter radius (0..7) — F0 contour smoothing. Higher = smoother but less expressive.
• Index rate (0.00..1.00) — faiss retrieval mix. Only takes effect when the loaded model ships with an .index.
• Denoise toggle, Bypass model toggle.

All knobs update live without restarting the engine.

Latency budget

Hardware	End-to-end
M1 / M2 (MPS)	~350–400 ms
M3 / M4 (MPS)	~300–370 ms
Intel / CPU	~400–500 ms

Most of the latency is in the 384 ms accumulation window. Lower --window-ms (or change in engine.py) trades quality for latency.

CLI

Voicebox also ships a CLI for headless / scripted use:

./venv/bin/python realtime_voice.py \
    --rvc-model <voice_name> \
    --pitch-shift 0 \
    --output-device <BlackHole device index>

Run --list-devices to see all CoreAudio devices.

Build from source

Requires macOS 13+, Apple Silicon recommended.

brew install python@3.10 cmake portaudio create-dmg
git clone https://github.com/<your-fork>/voicebox.git
cd voicebox
./setup.sh         # creates ./venv, installs deps, downloads HuBERT + rmvpe
./venv/bin/python ui.py

To produce a .app + .dmg:

./build_app.sh
# → dist/Voicebox.app (~1.6 GB) + dist/Voicebox.dmg (~1.1 GB)

The bundle uses ad-hoc signing only. For Apple Developer ID + notarization, edit Voicebox.spec and the codesign step in build_app.sh.

Architecture

ui.py
 └─ ui_window.MainWindow
     ├─ ui_setup.SetupScreen      ← system checks, models, downloads, log
     ├─ ui_pipeline.PipelineScreen ← routing, voice tuning, transport, telemetry
     └─ tray icon (close-to-tray)

engine.Engine
 ├─ audio_io.AudioIO  ← sounddevice streams + lock-protected ring buffers
 ├─ denoise.Denoiser  ← DeepFilterNet3 wrapper
 ├─ rvc.RVC           ← rvc-python primary, vendored fairseq fallback
 └─ pipeline.Pipeline ← sliding-window + equal-power crossfade

system_checks.py   ← homebrew / python / BlackHole / Multi-Output / base models / mic permission
models_manager.py  ← scan / drop / remove RVC voice folders
app_paths.py       ← user-data dir (works in source and frozen .app)

Specs and plans live in docs/superpowers/.

Troubleshooting

• No module named 'fairseq' — ./venv/bin/pip install rvc-python.
• aten::_fft_r2c not implemented for MPS — PYTORCH_ENABLE_MPS_FALLBACK=1 (already set in ui.py and realtime_voice.py).
• High latency / underruns — raise --blocksize (e.g. 960), disable Denoise, check Activity Monitor.
• No audio in Discord/Zoom — verify BlackHole 2ch is the input in the target app, and Voicebox's output device is BlackHole.
• Voice models list squashed — fixed in v0.4.1: list scrolls inside a fixed-height area.
• macOS prompts for mic on every Start — only happens once per process. If it repeats, the bundle identifier may have changed; rebuild the .app.
• App "frozen" right after Start — first RVC model load takes 10–25 s on cold start; the transport button shows ⏳ Loading models…. Subsequent starts are fast (Engine reuses loaded models).
• sounddevice import error / PortAudio not found — ./venv/bin/pip install --force-reinstall --no-binary :all: sounddevice or brew reinstall portaudio.

Privacy & permissions

• Voicebox processes audio entirely locally — nothing is uploaded.
• It requests microphone access the first time you start the pipeline.
• Models, base encoders, and settings live in:
  • From source: the repo dir under models/ and ~/.config/microphone/.
  • From the .app: ~/Library/Application Support/Voicebox/.

Acknowledgements

• DeepFilterNet — Rikorose et al.
• Retrieval-based-Voice-Conversion-WebUI / RVC-Project.
• rvc-python — Python wrapper for RVC inference.
• ContentVec — multilingual content encoder.
• BlackHole — Existential Audio's virtual audio loopback driver.
• qtawesome — Font Awesome / Material Design icons for Qt.

Author

Built by acrossoffwest. Contributions via PR welcome.

License

MIT — see LICENSE. Copyright © 2026 acrossoffwest.