Persona Voice replaces assistant speech locally, with near-real-time playback.
Persona Voice replaces the spoken output of ChatGPT, Codex, or Grok Bot with a voice you choose. Select the source application once; its conversation and controls stay untouched while conversion runs locally on your device. Output quality and timing vary with the hardware, source audio, and selected reference.
| Grok Bot | Codex |
|---|---|
Voice-changer-Grok-demo-40s-v3.mp4 |
demo.mp4 |
Download one model during first-run setup, then add the other later in Settings → Voice model. Existing installations keep their current model; new Apple Silicon installs recommend Chatterbox.
| Model | Released | Speech and performance |
|---|---|---|
| Seed-VC Tiny | 2024 | Best suited to Japanese and Chinese in our comparisons; 300 ms streaming blocks. Apple Silicon and x64 NVIDIA CUDA. |
| Chatterbox | 2025 | Recommended for English and other languages; English live-tested. 640 ms streaming blocks on Apple Silicon. |
On the same English recording, Chatterbox took 29% less processing time on our Apple M4 Pro. Quality varies by language and voice. See model comparison and sources for measurements, startup delay, storage requirements, and availability.
- Near-real-time conversion. Both models process speech in bounded blocks and streams converted audio as it becomes available. Actual latency varies by hardware and route.
- The original voice is replaced, not layered. Persona Voice suppresses the selected app's original playback and sends the converted voice to your speakers.
- One relay, multiple voice apps. Switch between ChatGPT (including Codex) and Grok Bot without changing models, voices, or the local conversion pipeline.
- Local inference. Once installed, conversion runs on your device. No voice API key is required.
- Voice presets and local references. The included catalog contains credited VOICEVOX identities and a small set of community/demo references. You can also add an authorized private reference of your own.
- Personalisation. Pick a bundled identity or pair an authorized reference with its own character scene.
- Private history controls. History is off by default. If enabled, only converted output can be stored, with six-hour cleanup by default and an immediate clear action.
ChatGPT / Codex or Grok Bot
│ voice output
▼
Persona Voice audio route
▼
Selected local model: Chatterbox / Seed-VC
│
▼
Speakers
Persona Voice waits until the selected app, local engine, and output route are ready before replacing the original playback. If the route cannot be established safely, conversion does not start.
Read the full architecture, native protocol, and engine contract.
Download the latest macOS, Windows, or Linux build from Releases. Windows setup links to the official VB-CABLE download; install it separately, restart Windows, and follow the in-app Volume Mixer step.
- Launch Persona Voice, choose ChatGPT or Grok Bot, complete system-audio setup, and download one voice model.
- Select a target voice and press Start voice.
- Open voice mode normally in the application you selected.
See Platform status for requirements and Troubleshooting if setup is blocked.
Requirements:
- Git, Bun 1.3.14, Node.js 22.12+, and
uv; - one qualified host profile: Apple Silicon macOS 14.2+ with MPS/MLX, x64 Linux with a supported NVIDIA CUDA driver, or x64 Windows build 20348+ with a supported NVIDIA CUDA driver;
- the platform native toolchain: Xcode Command Line Tools on macOS, a C++20 compiler plus
pkg-config/PipeWire development headers on Linux, or MSVC/CMake/Windows SDK on Windows; - Chatterbox space (macOS): approximately 4 GiB installed, 8 GiB free for setup. Seed-VC space: 2.5 GiB / 6 GiB free on macOS, 9 GiB / 15 GiB free on Windows, 11 GiB / 15 GiB free on Linux.
- Windows also requires the official VB-CABLE driver, installed separately from VB-Audio.
git clone --recurse-submodules https://github.com/miuuyy/ChatGPT-Persona-Voice.git
cd ChatGPT-Persona-Voice
bun install --frozen-lockfile
bun run devLinux source runs also require PipeWire and WirePlumber. See Development for platform setup, native build commands, and contributor verification.
| Platform | Availability | Requirements and current limits |
|---|---|---|
| Apple Silicon macOS 14.2+ | Preview package available | MPS/MLX; production signing/notarization and clean-machine qualification remain |
| Linux x64 + NVIDIA | Preview package available | CUDA 13.0, PipeWire, and WirePlumber; broader distribution coverage remains |
| Windows x64 + NVIDIA, build 20348+ | Preview package available | CUDA 13.0 and separately installed VB-CABLE; physical-host feedback is welcome |
| Other hosts | Unavailable | Unsupported |
See the detailed platform matrix and release gates.
The bundled catalog currently includes Shikoku Metan, Zundamon, Kasukabe Tsumugi, Meimei Himari, Kyushu Sora, WhiteCUL, Ouka Miko, Sayo, Haruka Nana, Nekotsuka Aru, Manbetsu Hanamaru, Kotoyomi Nia, a community JARVIS reference, and an unaffiliated Donald Trump demo likeness.
VOICEVOX samples are assembled from official showcase audio and retain their required credit. Community and public-figure references retain their own terms and must never be presented as authentic speech or endorsement. Use only voices you are authorized to use. See the voice manifest and the single third-party notice inventory.
- Raw captured PCM is not intentionally persisted or logged.
- History accepts only converted frames submitted to the output session.
- Voice replacement starts only after the local engine and audio route are ready.
- Settings, logs, models, references, and optional history remain in local workspace/application storage during use.
- On macOS, BlackHole and OBS are separate trust boundaries. When using the converted-only recording bus, mute audio from OBS macOS Screen Capture or it will record the original system stream as well.
Read Privacy, Security, and Troubleshooting before using sensitive audio.
bun run test
bun run typecheck
bun run build:renderer
bun run check
bun run smoke:engine| Document | Contents |
|---|---|
| Development | Setup, checks, native smokes, and contribution workflow |
| Architecture | Process boundaries, lifecycle, queues, and persistence |
| Platform matrix | Implemented paths and remaining release gates |
| Native protocol | CPV1 framing and bounded audio transport |
| Engine contract | CPVE lifecycle and Seed-VC profile |
| Model adapters | Rules for integrating another conversion backend |
| Release engineering | Artifact policy, signing, and publication gates |
Contributions are welcome within the current experimental scope. Start with CONTRIBUTING.md and follow the Code of Conduct.
Original launcher code is available under the MIT License. Seed-VC remains GPL-3.0, and model files, voice references, and dependencies retain their own licenses and terms. See Third-party notices.
Persona Voice is independent software and is not affiliated with or endorsed by OpenAI, xAI, or Cursor. ChatGPT, Codex, OpenAI, Grok, Grok Bot, xAI, and Cursor names and marks belong to their respective owners. This project does not bypass authentication, subscriptions, permissions, or access controls.