VoiceBox: The Open-Source AI Voice Studio With 46,000+ GitHub Stars — Clone, Dictate, and Give Your AI Agents a Voice

What Is VoiceBox and Why Developers Are Obsessed

If you've been following GitHub Trending lately, you've probably noticed VoiceBox climbing the charts with over 46,000 stars and counting. Created by Jamie Pine, this open-source AI voice studio is making serious waves in the developer community — and for good reason.

VoiceBox solves a fundamental problem: the voice I/O stack has been split between expensive cloud services. ElevenLabs dominates voice output (text-to-speech), while WisprFlow handles voice input (dictation). Both require subscriptions, send your data to the cloud, and lock you into their ecosystems.

VoiceBox does both, runs entirely on your machine, and costs nothing. It's the complete voice pipeline — cloning, generation, dictation, effects, and agent integration — in a single native app.

7 TTS Engines: Choose the Right Voice for Every Task

Unlike most TTS tools that offer one engine and hope it fits your use case, VoiceBox ships with seven different text-to-speech engines, each optimized for different scenarios:

Qwen3-TTS (0.6B / 1.7B)

The multilingual powerhouse. Supports 10 languages with high-quality voice cloning and natural-language delivery instructions. Tell it to "speak slowly" or "whisper" and it actually does.

// Example: Delivery control with Qwen3-TTS
const text = "Welcome to our tutorial. [pause] Let's begin with the basics.";
// Qwen3-TTS interprets delivery instructions naturally

Chatterbox Multilingual

The language coverage champion — 23 languages including Arabic, Japanese, Hindi, Swahili, Turkish, and more. If you're building for a global audience, this is your engine.

Chatterbox Turbo

The expressive one. This 350M model is the only engine that understands paralinguistic tags like [laugh], [sigh], [gasp], and [chuckle]. Perfect for conversational AI, podcasts, and character voices.

// Chatterbox Turbo understands emotion tags
"Well, that's unexpected! [laugh] I didn't see that coming. [sigh] Let me think about it..."

LuxTTS

The lightweight option. Only ~1GB VRAM, 48kHz output, and runs at 150x realtime on CPU. Great for quick iterations and resource-constrained environments.

HumeAI TADA (1B / 3B)

The long-form specialist. Generates coherent audio up to 700+ seconds with text-acoustic dual alignment. Ideal for audiobooks, long tutorials, and podcast episodes.

Kokoro

Tiny (82M model), fast, and comes with 50 curated preset voices. Perfect for prototyping and CPU-only setups.

Qwen CustomVoice

9 preset voices with natural-language delivery control — no reference audio required. Just describe how you want it to sound.

Voice Cloning: From 3 Seconds to Your Own Voice Model

VoiceBox's zero-shot voice cloning is where things get really interesting. Here's how it works:

1. Record or upload a few seconds of audio (3-10 seconds is enough)
2. Create a profile in the app — name it, add a description
3. Generate speech in that voice, in any of the supported languages

The quality is surprisingly good. Multi-sample support (uploading several audio clips of the same voice) improves cloning accuracy even further.

// Clone a voice and generate speech via API
const response = await fetch('http://localhost:3000/api/generate', {
  method: 'POST',
  headers: { 'Content-Type': 'application/json' },
  body: JSON.stringify({
    text: "Hello, this is my cloned voice speaking!",
    profile: "my-voice-profile",
    engine: "qwen3-tts"
  })
});

const audio = await response.blob();
// Use the audio blob in your app

Global Dictation: Type With Your Voice Anywhere

This is where VoiceBox goes beyond just TTS. The global dictation hotkey lets you dictate into any text field on your system — not just inside the app.

Hold your configured hotkey, speak, release. On macOS, the transcript automatically pastes into whatever text field has focus. It uses OpenAI Whisper for transcription (the same model behind most cloud dictation services), running locally on MLX (Apple Silicon) or PyTorch (CUDA/ROCm/CPU).

Available Whisper sizes: Base, Small, Medium, Large, and Turbo (~8x faster than Large with minimal quality loss).

LLM Refinement

Raw dictation often includes "umms," false starts, and stutters. VoiceBox includes an optional local LLM refinement step that cleans up your transcript before pasting — removing filler words, fixing self-corrections, and preserving technical terms.

MCP Integration: Give Your AI Agents a Voice

This is the feature that has developers most excited. VoiceBox includes a built-in MCP (Model Context Protocol) server that lets any MCP-aware AI agent speak to you.

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

Imagine: Claude Code tells you when a build finishes, Cursor narrates code reviews, or your custom agent speaks task completions — all in a voice you've cloned or chosen. No more staring at terminals waiting for output.

The MCP server supports both HTTP and stdio transports, so it works with Claude Code, Cursor, Windsurf, VS Code MCP, and any client that speaks the protocol.

Per-Agent Voice Binding

Pin different voices to different agents. Claude Code gets "Morgan," Cursor gets "Scarlett," your custom agent gets "Atlas." You always know which agent is talking without looking at the screen.

Post-Processing: Studio-Quality Audio Effects

VoiceBox includes 8 audio effects powered by Spotify's Pedalboard library:

• Pitch Shift — Up or down by 12 semitones
• Reverb — Room size, damping, wet/dry mix
• Delay — Echo with time, feedback, and mix controls
• Chorus / Flanger — Modulated delay for texture
• Compressor — Dynamic range compression
• Gain — Volume adjustment (-40 to +40 dB)
• High-Pass Filter — Remove low frequencies
• Low-Pass Filter — Remove high frequencies

Ships with 4 built-in presets (Robotic, Radio, Echo Chamber, Deep Voice) and supports custom presets. Effects can be assigned per-profile as defaults.

Stories Editor: Multi-Track Audio Composition

Need to create a podcast episode, a dialogue scene, or a narrated tutorial? The Stories editor is a multi-track timeline where you can:

• Arrange multiple voice clips on a timeline
• Trim and split audio inline
• Pin specific voice versions to each clip
• Auto-playback with synchronized playhead

Think of it as a simplified DAW (Digital Audio Workstation) specifically designed for voice content.

Real-World Example: Building a Voice-Enabled Documentation Bot

Let's put it all together. Here's how you could build a voice-enabled documentation assistant using VoiceBox's API:

// 1. Start VoiceBox (it runs as a local server)
// 2. Use the REST API from your app

// Transcribe user's voice question
const transcribe = await fetch('http://localhost:3000/api/transcribe', {
  method: 'POST',
  body: audioBlob
});
const question = await transcribe.json();

// 3. Get answer from your AI/LLM
const answer = await getDocumentationAnswer(question.text);

// 4. Speak the answer back using a cloned voice
await fetch('http://localhost:3000/api/speak', {
  method: 'POST',
  headers: { 'Content-Type': 'application/json' },
  body: JSON.stringify({
    text: answer,
    profile: "support-agent",
    engine: "qwen3-tts"
  })
});

The entire pipeline runs locally. No cloud API costs, no data leaving your network, no rate limits.

Key Benefits

• Complete Privacy — Models, voice data, and audio captures never leave your machine
• Zero Cost — Free and open-source, no subscriptions or usage fees
• Native Performance — Built with Tauri (Rust), not Electron. Fast and lightweight
• 7 TTS Engines — Pick the right engine for each use case
• 23 Languages — From English to Arabic, Japanese, Hindi, Turkish, and more
• Voice Cloning — Zero-shot cloning from a few seconds of audio
• Global Dictation — Dictate into any app with a system-wide hotkey
• MCP Integration — Give Claude Code, Cursor, and other agents a voice
• Post-Processing Effects — 8 studio-quality audio effects
• Unlimited Length — Auto-chunking with crossfade for long-form content
• Cross-Platform — macOS (Apple Silicon + Intel), Windows, Linux, Docker
• API-First — REST API + MCP server for easy integration

Getting Started

VoiceBox is available for all major platforms:

GPU acceleration is supported: MLX/Metal on macOS, CUDA on Windows/Linux, AMD ROCm, and Intel Arc. CPU-only mode works everywhere.

Frequently Asked Questions

Is VoiceBox really free?

Yes, completely free and open-source under the MIT license. No subscriptions, no usage fees, no cloud costs. All models run locally on your machine.

How good is the voice cloning quality?

Surprisingly good, especially with Qwen3-TTS and Chatterbox Multilingual. A 3-10 second reference sample is enough for recognizable cloning. Multi-sample profiles (uploading several clips) produce even better results.

Does it work without an internet connection?

Yes! VoiceBox is local-first. Once models are downloaded on first use, everything works offline. Your voice data never touches a cloud server.

What GPU do I need?

It depends on the engine. LuxTTS runs on ~1GB VRAM. Kokoro is 82M and works on CPU. Larger models like Qwen3-TTS 1.7B benefit from more VRAM but also work on CPU (just slower). Apple Silicon users get MLX acceleration out of the box.

Can I use VoiceBox in my commercial projects?

VoiceBox itself is MIT-licensed, so yes. However, be aware that some TTS model weights may have their own licenses. Check the individual engine licenses for commercial voice generation use.

How does the MCP integration work?

VoiceBox runs an MCP server that exposes a voicebox.speak tool. Any MCP-aware agent (Claude Code, Cursor, Cline, etc.) can call this tool to generate speech. Install it as a URL in your agent's MCP settings, or use the bundled voicebox-mcp binary for stdio transport.

What's the maximum text length?

50,000 characters per generation. VoiceBox uses smart auto-chunking (configurable 100-5,000 chars) with crossfade (0-200ms) to produce seamless long-form audio.

Does dictation work in every app?

On macOS, yes — the global hotkey triggers Whisper transcription and the result is accessibility-verified auto-pasted into whatever text field has focus. On Windows/Linux, the in-app mic works everywhere within VoiceBox.

Ready to level up your development skills? Explore interactive courses at CoddyKit — learn coding at your own pace with hands-on exercises. 🚀