The AI voice landscape has long been dominated by a handful of cloud-based services — ElevenLabs being the most prominent — where you pay per character, trust your audio to a third-party server, and accept their rate limits and pricing changes.
But that's changing.
Enter Voicebox by Jamie Pine — an open-source, self-hosted AI voice studio that puts everything on your machine. No subscriptions, no API keys, no data leaving your computer. And with over 1.3 million downloads and roughly 35,000 GitHub stars, it's clear the world was ready for this.
What Is Voicebox?
Voicebox is a desktop application (built with Tauri, React, and FastAPI) that turns your computer into a full-featured AI voice studio. It runs on macOS, Windows, and Linux, and packs voice cloning, text-to-speech generation, system-wide dictation, and even an MCP server for AI agents — all local.
Think of it as the Ollama of voice AI. Just as Ollama brought large language models to your desktop, Voicebox brings professional voice capabilities to your machine, free and private.
Voice Cloning From Just 3 Seconds
Voice cloning is the headline feature, and it's remarkably simple. You can feed it audio in three ways:
- Upload a clip — drag and drop WAV, MP3, FLAC, or WebM files
- Record from microphone — live waveform preview, up to 30 seconds
- System audio capture — clone a voice from a YouTube video, podcast, or any app playing audio
From as little as three seconds of audio, Voicebox creates a voice profile you can use across all its engines. Zero-shot cloning means no lengthy training process — just supply a sample and generate.
Seven TTS Engines, One Interface
Voicebox doesn't lock you into a single model. It ships with seven TTS engines, each with different strengths:
This breadth means you pick the right tool for the job. Need fast iterations at high quality? LuxTTS. Need expressive speech with emotional control? Chatterbox or TADA. Working on a low-resource machine? Kokoro runs on practically anything.
For transcription, Voicebox bundles Whisper and Whisper Turbo (99 languages, the latter 8x faster), plus the Qwen3 language model powers transcript cleanup and persona replies.
System-Wide Dictation: Your AI Typing Assistant
One of the most practical features is system-wide dictation. With a global hotkey (Cmd+Opt on macOS, Ctrl+Alt on Windows) you can speak into any app and have the transcript appear at your cursor. No more wrestling with speech-to-text browser extensions — it works everywhere, from code editors to chat apps.
This is the same experience WisprFlow charges for, built into a free open-source application.
MCP Server: Give Every AI Agent a Voice
If you use MCP-aware agents like Claude Code, Cursor, or Cline, Voicebox exposes an MCP server with a single voicebox.speak tool call. Any agent can talk to you through a voice you've cloned — complete with per-agent voice profiles.
You can have Claude Code speak in one voice, Cursor in another, and know which agent is talking without looking at the screen. Every agent-initiated speech surfaces a visible pill, so there's no silent background TTS.
REST API and WebSocket for Developers
Voicebox exposes a local REST API and WebSocket endpoints for every engine you download. This means:
- Games — generate NPC dialogue on the fly with your characters' voices
- Apps — build voice-enabled applications without paying per-character fees
- Scripts — batch-produce audiobook chapters, automate podcast intros, wire it to your Stream Deck
No API keys, no rate limits, no per-character fees. Just a localhost URL.
Personas: Voices With Character
A particularly creative feature is the Persona system. You can give any voice profile a free-form personality description ("1940s noir detective. World-weary, cynical..."), then:
- Rewrite — restate your text in that character's voice while preserving the ideas
- Compose — let the character improvise original lines from scratch
This is gold for game dialogue, dubbing, long-form narration, and any creative project where consistent character voice matters.
Privacy and Data Sovereignty
Everything runs locally. Voicebox downloads models to your machine and keeps all processing on-device. No audio clips are sent to a cloud server, no voice profiles leave your computer, and there are no telemetry calls sending usage data home.
In an era where every AI tool seems to want a slice of your data, Voicebox's commitment to local-first operation is genuinely refreshing.
How It Stacks Up Against ElevenLabs
ElevenLabs still wins on sheer voice quality polish and supported language count. But Voicebox competes fiercely on flexibility, privacy, and cost — and for many use cases, especially developer workflows, it's already the better choice.
Getting Started
Voicebox is available now at voicebox.sh or on GitHub. Download the app for your platform, pick the engines you want, and you're generating speech within minutes.
The project also has a Solana token ($VOICEBOX) for those who want to support it, but the application itself remains free and open-source forever — no token required.
The Bottom Line
Voicebox represents a shift in how we think about AI voice tools. By bringing everything local, embracing open-source principles, and supporting a broad ecosystem of TTS engines, it empowers users in a way that walled-garden services never can.
If you've been hesitating on voice AI because of cost, privacy concerns, or vendor lock-in, Voicebox is the answer. It's the Ollama of voice — and it's only getting better.
Continue reading: For a deeper technical dive into setting up Voicebox for production use, check out the official documentation at voicebox.sh/docs.