VoiceBox: The Open-Source AI Voice Studio With 43,000+ GitHub Stars That Lets You Clone Voices, Dictate, and Talk to AI Agents

If you've ever wanted to clone a voice, generate natural-sounding speech, or give your AI coding assistant a human voice — all without sending a single byte to the cloud — VoiceBox is the project you've been waiting for.

Created by Jamie Pine, VoiceBox has exploded onto the open-source scene, accumulating over 43,000 GitHub stars and gaining 629 new stars in a single day. It's not just another text-to-speech tool — it's a complete voice I/O stack that runs locally on your machine.

Think of it this way: ElevenLabs handles voice output. WisprFlow handles voice input. VoiceBox does both, bridges them with a bundled local LLM for refinement and per-profile personas, and never lets your voice data leave your computer.

Why VoiceBox Matters for Developers

The voice AI space has been dominated by expensive cloud services. ElevenLabs charges per character for TTS. WisprFlow requires a subscription for dictation. Both send your voice data to remote servers.

VoiceBox flips this model entirely:

• Complete privacy — models, voice data, and audio captures never leave your machine
• Zero cost — free and open-source, forever
• Native performance — built with Tauri (Rust), not Electron bloat
• API-first design — REST API plus a built-in MCP server for seamless integration
• Runs everywhere — macOS (Apple Silicon with MLX/Metal), Windows (CUDA), Linux, AMD ROCm, Intel Arc, and Docker

For developers building AI-powered applications, this means you can integrate voice cloning, speech generation, and dictation without worrying about API costs, rate limits, or data privacy compliance.

Seven TTS Engines: Choose the Right Voice for Every Task

Unlike single-engine solutions, VoiceBox ships with seven different TTS engines, each optimized for different use cases. You can switch engines per generation, giving you maximum flexibility:

Engine	Languages	Best For
Qwen3-TTS (0.6B / 1.7B)	10	High-quality multilingual cloning with delivery instructions ("speak slowly", "whisper")
LuxTTS	English	Lightweight (~1GB VRAM), 48kHz output, 150x realtime on CPU
Chatterbox Multilingual	23	Broadest language coverage — Arabic, Japanese, Hindi, Turkish, and more
Chatterbox Turbo	English	Fast 350M model with paralinguistic emotion tags like [laugh], [sigh]
HumeAI TADA (1B / 3B)	10	700+ seconds of coherent audio, text-acoustic dual alignment
Kokoro	8	50 curated preset voices, tiny 82M model, fast CPU inference
Qwen CustomVoice	10	9 preset voices with natural-language delivery control — no reference audio needed

This multi-engine approach means you're never locked into one model's limitations. Need fast English speech for a quick notification? Use LuxTTS. Creating a multilingual podcast? Switch to Chatterbox Multilingual. Want expressive speech with laughs and sighs? Chatterbox Turbo has you covered.

Voice Cloning That Actually Works

VoiceBox's voice cloning is where the magic happens. Unlike cloud services that require minutes of audio, VoiceBox can clone a voice from just a few seconds of reference audio.

How Voice Cloning Works

1. Create a profile — upload an audio file or record directly in the app
2. Multi-sample support — add multiple samples for higher quality cloning
3. Zero-shot cloning — no fine-tuning required, instant results
4. Import/export profiles — share cloned voices or back them up

Each voice profile can include a description, language tags, and a default effects chain. You can organize profiles, star favorites, and track the provenance of every generation.

Voice Personalities: Beyond Just Sound

VoiceBox goes beyond acoustic cloning with voice personalities. Attach a free-form personality to any voice profile — who this voice is, how they speak, what they care about. A bundled local LLM (Qwen3) powers two actions:

• Compose — generate fresh in-character lines with a shuffle button
• Speak in Character — rewrite your input text to match the personality's style

This means your cloned voice doesn't just sound like someone — it speaks like them.

Global Dictation: Type With Your Voice Anywhere

The other half of VoiceBox's voice I/O loop is global dictation. Hold a hotkey anywhere on your system, speak, release — and on macOS, the transcript pastes straight into whatever text field you're focused on.

Dictation Features

• Configurable chord bindings — hold-to-speak and tap-to-toggle modes, each rebindable
• Target-aware paste (macOS) — accessibility-verified injection into the focused text field with atomic clipboard save/restore
• In-app mic button — available on every VoiceBox text field
• LLM refinement — optional cleanup of ums, stutters, and false starts before paste
• Whisper-based STT — supports Base, Small, Medium, Large, and Turbo models

The Turbo Whisper model is approximately 8x faster than Whisper Large with minimal quality loss, making real-time dictation practical even on modest hardware.

The Captures Tab: Your Voice History

Every dictation session, in-app recording, and uploaded audio file lands in the Captures tab — original audio paired with transcript, always preserved. You can:

• Replay, re-transcribe with different Whisper sizes, or refine with the local LLM
• Edit transcripts inline and save on blur
• Play captures as any voice profile with one click
• Promote captures to voice samples for future cloning

MCP Integration: Give AI Agents a Voice

This is where VoiceBox becomes truly revolutionary for developers. With MCP (Model Context Protocol) integration, any MCP-aware AI agent can speak to you in a voice you've cloned.

// In any MCP-aware agent (Claude Code, Cursor, Cline):
await voicebox.speak({
  text: "Deploy complete. All tests passing.",
  profile: "Morgan",
});

One tool call, and your AI assistant has a voice. Task completions, questions, notifications — all spoken in a voice you own.

MCP Features

• Bidirectional pill overlay — recording, transcribing, refining, and speaking all surface in the same OS-level overlay
• Per-agent voice binding — pin Claude Code to one voice, Cursor to another, so you can tell which agent is talking without looking
• Always visible — no silent background TTS; every agent-initiated speak surfaces the pill with the voice profile name
• HTTP + stdio transports — install as a URL in Claude Code, Cursor, Windsurf, or VS Code MCP, or use the bundled voicebox-mcp binary for stdio-only clients

For developers building AI-powered applications, this opens up entirely new interaction paradigms. Your coding agent doesn't just print to a terminal — it tells you what's happening, in a voice you've chosen.

Post-Processing Effects: Professional Audio Without the DAW

VoiceBox ships with eight audio effects powered by Spotify's pedalboard library. Apply effects after generation, preview in real time, and build reusable presets:

Effect	Description
Pitch Shift	Up or down by up to 12 semitones
Reverb	Configurable room size, damping, wet/dry mix
Delay	Echo with adjustable time, feedback, and mix
Chorus / Flanger	Modulated delay for metallic or lush textures
Compressor	Dynamic range compression for consistent volume
Gain	Volume adjustment (-40 to +40 dB)
High-Pass Filter	Remove low frequencies (reduce rumble)
Low-Pass Filter	Remove high frequencies (reduce hiss)

Four built-in presets (Robotic, Radio, Echo Chamber, Deep Voice) get you started, and you can create custom presets. Effects can be assigned per-profile as defaults, so every generation from a specific voice automatically applies your preferred processing chain.

Stories Editor: Multi-Track Audio Composition

For podcasters, content creators, and anyone building multi-voice narratives, VoiceBox includes a Stories editor — a multi-track timeline for conversations and narratives.

• Multi-track composition with drag-and-drop
• Inline audio trimming and splitting
• Auto-playback with synchronized playhead
• Version pinning per track clip

Create conversations between cloned voices, build audio dramas, or produce educational content with multiple narrators — all within VoiceBox.

Real-World Example: Building a Voice-Enabled AI Assistant

Let's walk through a practical scenario. Say you're building a personal AI assistant that runs locally on your machine. You want it to:

1. Listen to your voice commands
2. Process them with a local LLM
3. Respond in a cloned voice that sounds like your favorite narrator

Step 1: Clone Your Preferred Voice

Download a few seconds of audio from your favorite audiobook narrator or podcast host. Create a VoiceBox profile, upload the sample, and you now have a cloned voice profile called "Narrator."

Step 2: Set Up Dictation

Configure VoiceBox's global hotkey (say, Cmd+Shift+Space). Now you can hold that key anywhere on your system, speak your command, and release. VoiceBox transcribes your speech using Whisper and pastes it into your AI assistant's input field.

Step 3: Connect Your AI Agent via MCP

In your AI assistant's MCP configuration, install VoiceBox's MCP server. Bind the "Narrator" voice profile to your assistant. Now when your assistant generates a response, it calls:

await voicebox.speak({
  text: responseText,
  profile: "Narrator",
});

Your assistant speaks the response in the cloned voice, with the pill overlay showing you exactly what's being said. The entire loop — voice input, AI processing, voice output — runs locally with zero cloud dependency.

Step 4: Add Personality

Attach a personality to the "Narrator" profile: "Warm, knowledgeable, speaks in complete sentences, occasionally uses metaphors." Enable "Speak in Character" mode, and your assistant's responses are rewritten to match the personality before being spoken.

The result? A fully voice-enabled AI assistant that sounds and speaks exactly how you want it to, running entirely on your hardware.

Key Benefits for Developers and Creators

Getting Started with VoiceBox

VoiceBox is available for macOS (Apple Silicon and Intel), Windows, and Docker. Linux users can build from source.

Quick Install

macOS (Apple Silicon):

# Download the DMG from voicebox.sh/download/mac-arm
# Or use Homebrew (when available)
brew install --cask voicebox

Windows:

# Download the MSI from voicebox.sh/download/windows

Docker:

docker compose up

After installation, VoiceBox walks you through first-run permissions (macOS Accessibility and Input Monitoring for dictation). From there:

1. Create your first voice profile from an audio sample
2. Generate speech in your cloned voice
3. Configure the global dictation hotkey
4. Install the MCP server for AI agent integration

The Roadmap: What's Coming Next

VoiceBox's development is active, with planned features including:

• More STT engines — Parakeet v3, Qwen3-ASR for even faster/more accurate transcription
• Linux pre-built binaries — currently build-from-source only
• Additional TTS engines — community contributions welcome
• Enhanced MCP capabilities — deeper agent integration

Frequently Asked Questions

Is VoiceBox really free?

Yes, VoiceBox is completely free and open-source under the MIT license. There are no premium tiers, no per-character charges, and no subscription fees. You own your voice data, your cloned profiles, and everything generated.

How much audio do I need to clone a voice?

VoiceBox uses zero-shot voice cloning, which means you can create a voice profile from just a few seconds of reference audio. However, multi-sample support allows you to add more samples for higher quality results. For best results, use clear audio without background noise.

Does VoiceBox work offline?

Absolutely. VoiceBox is designed to be local-first. All models (TTS engines, Whisper STT, Qwen3 LLM for refinement) download to your machine and run entirely offline. Once installed, you never need an internet connection to generate speech or transcribe audio.

Can I use VoiceBox for commercial projects?

Yes. VoiceBox is MIT-licensed, which permits commercial use. You can integrate it into commercial products, use cloned voices in commercial content, and build paid services on top of the REST API or MCP server.

What hardware do I need to run VoiceBox?

VoiceBox supports a wide range of hardware. On Apple Silicon, it uses MLX/Metal acceleration. On NVIDIA GPUs, it uses CUDA. AMD GPUs use ROCm, and Intel Arc uses DirectML. For CPU-only usage, LuxTTS runs at 150x realtime with just ~1GB VRAM. The Kokoro engine (82M parameters) runs efficiently on CPU alone.

How does VoiceBox compare to ElevenLabs?

VoiceBox offers similar voice cloning quality to ElevenLabs but with key advantages: it's free, runs locally (no API costs or rate limits), keeps your data private, and includes dictation features that ElevenLabs doesn't offer. However, ElevenLabs may have a slight edge in some voice quality metrics due to larger training datasets. The trade-off is cost, privacy, and control vs. marginal quality differences.

Can I use VoiceBox with Claude Code, Cursor, or Cline?

Yes! VoiceBox includes a built-in MCP server that integrates with any MCP-aware agent. Install it as a URL in Claude Code, Cursor, Windsurf, or VS Code MCP, or use the bundled voicebox-mcp binary for stdio-only clients. You can bind specific voice profiles to specific agents, so each one speaks in a different voice.

What languages does VoiceBox support?

Language support varies by engine. Chatterbox Multilingual covers 23 languages including Arabic, Danish, Finnish, Greek, Hebrew, Hindi, Malay, Norwegian, Polish, Swahili, Swedish, Turkish, and more. Qwen3-TTS covers 10 major languages. Kokoro supports 8 languages. You can switch engines per generation to access the language you need.