Double Chin

Your voice, on script.

Double Chin is a local voice-cloning application. Give it a few minutes of someone's voice and a text script, and it reads the script aloud in that voice — entirely on your machine. No cloud, no account, no audio leaving the laptop. One command opens the app: pick a voice, paste a script, hit Generate, and watch it synthesize chunk by chunk with a live speaker-similarity verdict on every take.

Watch a real session, script to spoken take:

double-chin-studio-demo.mp4

Hear it

Real clips of the owner's voice, cloned by Double Chin from a fine-tuned local model — nothing here is a recording. GitHub doesn't support inline audio players, so each clip below is a static-frame clip carrying the real generated audio track — press play, no download:

example-1.mp4

example-2.mp4

example-3.mp4

example-4.mp4

example-5.mp4

Each one verifies against a held-out reference clip the model never saw. The underlying voice — the reference audio, the fine-tuned adapter — never leaves this machine and isn't in this repo; only these generated clips are. Raw files: demo/owner_finetuned_demo.wav, demo/examples/. Full method and numbers: .goal/ledger.md (gate PASS 0.822, near-indistinguishable, threshold 0.70).

Requirements

• macOS on Apple Silicon (tested: M5 Pro, macOS 26) — Linux/CUDA and CPU also work, slower on CPU
• Python 3.12 (a managed one is fine; uv handles it)
• ~6 GB disk for model weights (downloaded once from Hugging Face on first run)

Setup

git clone <this-repo>&&cd double-chin
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -e ".[dev]"source .venv/bin/activate
double-chin doctor # checks device, deps, disk — everything should PASS
double-chin studio # starts http://127.0.0.1:8787 and opens your browser

Clone your own voice

Record the 79-take corpus (~50 minutes), then:

double-chin enroll yourvoice path/to/recordings/
double-chin train yourvoice path/to/recordings/ --manifest path/to/recordings/manifest.tsv
double-chin say "Hello, world." --voice yourvoice -o out.wav --verify

Every enrolled voice lives only in ~/.double-chin/voices/; deleting that folder revokes it entirely. Don't clone a voice you don't have the right to clone — every output carries an inaudible Perth watermark, but it's a provenance signal, not a control.

Licences

Double Chin: MIT. Engine (Chatterbox TTS): MIT. Verifier (resemblyzer): Apache 2.0.