Voicebox is an open-source desktop voice studio that clones voices, generates speech and turns your dictation into text entirely on your own computer. It is aimed at anyone who currently pays a subscription for a hosted voice service — video makers, podcasters, writers who dictate instead of typing — and at developers who want a voice layer their own scripts and agents can call without sending audio to someone else's servers.
What it does
The project describes itself as the full voice input/output stack running locally, and that covers three jobs that are usually three separate products:
- Voice cloning. Feed it a short sample — the video shows seconds of audio — and it builds a voice you can then speak text in.
- Speech generation. Type or paste text and get audio out, including expressive touches the project calls out specifically: laughter, sighs and whispers rather than a flat read.
- Dictation anywhere. A single global hotkey drops transcribed speech into whatever application currently has focus, which is the Wispr Flow-style workflow rather than a transcription window you copy out of.
On top of those, the README points at an API, and the channel's demo uses it to give a coding agent a spoken reply in a cloned voice. That is the part that makes Voicebox more than a media tool: the same engines that produce a narration track can be wired into an assistant loop.
How it works
Voicebox is a TypeScript application distributed as a downloadable release, with a landing site at voicebox.sh and separate documentation at docs.voicebox.sh. The interesting design decision is that it does not bet on one model. The video counts seven different AI engines available inside the app, all running on the local machine, so a user can swap the model behind cloning or transcription instead of accepting whatever a vendor ships.
The repository's own topic list names the pieces: Whisper on the recognition side, Qwen3-TTS on the synthesis side, plus CUDA and MLX. Those last two are the acceleration paths — one for NVIDIA GPUs, one for Apple's machine-learning framework — which is how a stack of local models stays fast enough to feel like dictation rather than batch processing. Nothing here is a thin client around a hosted endpoint; the models are the thing you install.
Getting started
The README's own navigation is the shortest route: a Download section pointing at the GitHub releases page, a Features section, an API section and a linked troubleshooting page in the docs. Voicebox ships built binaries and tracks total downloads on its release badge, so the normal path is to install the app rather than to build from source.
Two things are worth checking before you commit an evening to it. First, hardware: local models want a GPU or Apple silicon to be pleasant, and the CUDA and MLX topics tell you which two paths the project has invested in. Second, disk and model setup — seven engines are not seven small files, and the troubleshooting page exists for a reason. The licence is MIT, so both personal and commercial use are straightforward, and the source is open if you want to read what the app does with your recordings.
When to use it / when not
Use it when the audio itself is sensitive or the volume is high. Client recordings, internal meetings, medical or legal dictation, a personal voice you do not want stored on a third party's disk — those are cases where local processing is not a preference but a requirement. It also makes sense when a subscription no longer matches usage: a fixed one-off install beats per-character billing once you generate a lot.
Skip it when you have no capable hardware, when you need voice on a phone or inside a browser tab, or when you want a managed service with a support contract and an uptime guarantee behind it. A self-hosted stack of seven models is something you own, which also means something you maintain.
And the obvious caution: voice cloning from seconds of audio is a capability with real misuse potential. Clone voices you have permission to clone.
Alternatives
The comparison the project invites is with the hosted tools it set out to displace — ElevenLabs for cloning and speech generation, Wispr Flow for hotkey dictation. Those remain easier: nothing to install, nothing to accelerate, and they work on any machine. Voicebox trades that convenience for privacy, no recurring cost and the ability to change the model underneath. You can also run Whisper on its own for transcription, but you get one job rather than a studio with cloning, generation, dictation and an API in one place.
Take Voicebox seriously if you produce voice work regularly and have the machine to run it. Fifty-four thousand stars within a year of its first commit say the demand for a local alternative to subscription voice AI is real, and the combination of expressive cloned speech, system-wide dictation and a callable API is unusual to find in one MIT-licensed application. Anyone building an assistant that should talk back in a specific voice should read the API section first.