An open-source (MIT-licensed) AI voice studio built by Jamie Pine. It combines the “output” side — cloning a voice from a few seconds of audio and having it read text aloud — with the “input” side — a global hotkey that lets you dictate into any app — in a single desktop application. It is designed to replace both ElevenLabs and WisprFlow at once, and because the models and voice data all run on your own machine, no audio ever leaves it. The app is a native Tauri (Rust) build available for macOS, Windows, and Linux.
Key Features
- Seven switchable TTS engines: Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, and Kokoro can be selected per generation. Use Chatterbox Multilingual for broad language coverage (23 languages), LuxTTS when you want something lightweight, or Kokoro for preset voices.
- Voice cloning from a short sample: Zero-shot cloning works from a single reference recording. Kokoro and Qwen CustomVoice also ship more than 50 preset voices, so you can start without supplying a voice of your own.
- Dictation into any application: Hold the global hotkey and speak, and on macOS the transcript is pasted straight into the focused text field. Transcription runs OpenAI Whisper (including Turbo) locally. Every recording and transcript is kept in the Captures tab, where it can be re-transcribed or promoted into a voice profile.
- A built-in MCP server that gives agents a voice: The app ships a Model Context Protocol server, so agents such as Claude Code, Cursor, and Windsurf can call
voicebox.speakand talk back to you in a voice you cloned. A REST API (POST /speak,/generate,/transcribe, and more) is exposed on the same port. - Stories editor and an effects pipeline: A multi-track timeline editor lets you assemble multi-speaker conversations and podcasts. After generation you can apply eight audio effects — pitch shift, reverb, delay, compression, and others, powered by Spotify’s pedalboard — and save them as reusable presets.
- Automatic chunking for long text: Text is split at sentence boundaries, generated chunk by chunk, and crossfaded back together. It handles up to 50,000 characters, which is enough for full articles or book chapters.
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Open-source build (the only distribution) | $0 | Every feature, free of charge. The MIT license also permits commercial use, modification, and redistribution. |
Because all inference runs on your own machine, there is no service fee — the real cost is the hardware, meaning GPU and memory. macOS (Apple Silicon) is accelerated via MLX/Metal, NVIDIA GPUs on Windows via CUDA, AMD via ROCm, and Intel Arc via IPEX/XPU. CPU-only operation works everywhere, just more slowly.
Pricing information is current as of August 2026. Please check the official site for the latest details.
Pros & Cons
✅ Pros
- Neither voice data nor models leave your machine, which makes it easier to work with recordings and personal voices that cannot be uploaded elsewhere.
- There is no usage-based billing or character cap, so regenerating long narration costs nothing extra.
- Speech output (TTS) and speech input (dictation and transcription) live in one app, removing the need to switch tools.
- MCP and a REST API are included by default, making it straightforward to wire into agents and your own scripts.
- It is MIT-licensed open source, so adding engines or modifying the app is explicitly allowed.
⚠️ Cons
- You have to download models and configure your GPU; it is not a cloud service you can simply open and use.
- Generation speed depends on your own hardware, and some engines are not fast enough to be practical on CPU alone.
- Automatic paste for dictation currently works on macOS only; Windows and Linux are still on the roadmap.
- No prebuilt Linux binaries are provided, so Linux users need to build from source.
- Only Chatterbox Turbo interprets paralinguistic tags such as
[laugh]and[sigh]; other engines read them aloud as literal text.
Comparison with Similar Services
| Criteria | Voicebox | ElevenLabs | WisprFlow | OpenAI Audio API |
|---|---|---|---|---|
| Where it runs | Local | Cloud | Cloud | Cloud |
| Speech output (TTS) | Yes (7 engines) | Yes | No | Yes |
| Speech input (transcription) | Yes (Whisper) | Yes | Yes (core feature) | Yes |
| Voice cloning | Yes (zero-shot) | Yes | No | Limited |
| Pricing model | Free and open source | Subscription | Subscription | Usage-based |
| License | MIT | Proprietary | Proprietary | Proprietary |
Who Is It For
- People who need narration or transcription in environments where audio cannot be uploaded to an outside service
- Creators producing narration or podcasts at volume who would rather not watch a usage meter
- Developers who want agents like Claude Code or Cursor to announce finished work out loud
- Anyone who speaks faster than they type and wants voice input available in every app
- Open-source-minded users who want to add or modify engines themselves
Summary
Voicebox pulls the two halves of voice I/O — output, previously ElevenLabs’ territory, and input, previously WisprFlow’s — into a single locally run desktop app. It is free under the MIT license, carries no usage-based billing, and can give AI agents a voice through its MCP server. In exchange, you handle model downloads and GPU setup yourself, so it cannot match a cloud service for convenience. If privacy and cost matter most to you, it is a strong first choice: start with a lightweight engine such as LuxTTS or Kokoro on your own machine, then move up to the heavier engines as needed.