Every tier of AI text-to-speech in one native Windows app. Fully local. Free forever.
Voice cloning from a few seconds of audio, ElevenLabs-grade narration, instant drafts, multi-voice dialog scripts, batch queues, and studio effects — with nothing ever leaving your machine. No account, no subscription, no word limits, no cloud.
| Tier | Engine | Voice cloning | Quality | Speed |
|---|---|---|---|---|
| 3 | Chatterbox | ✓ | Beat ElevenLabs 65%–24% in blind tests | ~3s per word |
| 3 | F5-TTS | ✓ | Research-grade cloning specialist | ~4s per word |
| 2 | Kokoro | presets (50+) | Solid drafts | near-instant |
Slow is fine — queue a video script, walk away, come back to studio narration that cost $0.
Option A — release installer (recommended)
- Install Python 3.12 (check "Add to PATH"), then:
pip install -r https://raw.githubusercontent.com/The-Berin/FreeVoice/main/requirements.txt - Grab
FreeVoice-Studio-Setup.exefrom Releases and run it. - Launch FreeVoice Studio. First generation with each engine downloads its model automatically (Chatterbox ~1 GB, F5 ~1.4 GB, Kokoro ~350 MB).
Option B — from source
git clone https://github.com/The-Berin/FreeVoice
cd FreeVoice
pip install -r requirements.txt
dotnet build src\FreeVoiceStudio -c Release
& "src\FreeVoiceStudio\bin\Release\net8.0-windows\FreeVoice Studio.exe"The app supervises the Python engine server itself — start the app, everything else is automatic. ffmpeg on PATH is needed for MP3 export (WAV works without it).
- Studio — paste a script, pick an engine + voice, hit Generate. Live word count and render-time estimate. Jobs queue up and run in order with progress and ETA; play results inline.
- Multi-voice scripts — start a line with
[VoiceName]to switch speakers:[Woody] Somebody's poisoned the water hole! [Michael] Calm down, it's just a kokoro preset. - Voices — clone anyone from 7–20 seconds of clean speech (no music/noise). Add an optional transcript of the sample for better F5 accuracy. Cloned voices work with both Tier 3 engines.
- Delivery controls — Emotion & Pace (Chatterbox), Speed (Kokoro/F5), Quality steps (F5).
- Effects — Deep Voice, Radio, Echo Chamber, Robotic (ported from VoiceBox).
- Clean audio — spectral denoise (kills synthesis hiss) + loudness normalize to −16 LUFS (YouTube standard). On by default.
- Library — every generation, playable and deletable, stored in
output/.
The engine server listens on http://127.0.0.1:7899 while the app runs (or run it standalone: python server.py).
POST /api/generate {script, engine, voice, params{...}, effect, clean, format, title}
GET /api/state engines, voices, job queue with progress, outputs
POST /api/voices multipart: file, name, transcript
Full endpoint reference: docs/API.md. There's also a one-shot CLI:
python freevoice.py --file script.txt --out output\video.mp3 --voice voices\Woody.wav- docs/SETUP.md — detailed install, model storage, updating
- docs/API.md — REST endpoints for pipelines
- docs/ENGINES.md — engine details, licenses, tuning tips, troubleshooting
- 100% offline after models download. Voice data never leaves the machine.
- F5-TTS weights are CC-BY-NC (non-commercial license) — know what that means for your use.
- Chatterbox is MIT. Kokoro is Apache 2.0.
- CPU-only by design; a CUDA GPU would speed things up but is not required.