Voicebox

🚀 Description

jamiepine/voicebox — local-first AI voice studio. Free and open-source alternative to ElevenLabs (TTS) and WisprFlow (STT) in one app. Clone any voice from a few seconds of audio, generate speech in 23 languages across 7 TTS engines, dictate into any text field with a global hotkey, give any MCP-aware agent a voice you’ve cloned.

Built with Tauri (Rust), not Electron. Privacy by default — models, voice data, and captures never leave the machine.

🧩 Features

• Complete privacy — runs entirely locally
• 7 TTS engines — Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, Kokoro
• Voice cloning — zero-shot from a reference sample, or 50+ preset voices
• 23 languages — incl. Polish, Arabic, Japanese, Hindi, Swahili
• Post-processing — pitch shift, reverb, delay, chorus, compressor, filters (Spotify pedalboard)
• Expressive speech — paralinguistic tags [laugh] [sigh] [gasp] via Chatterbox Turbo
• Unlimited length — auto-chunking + crossfade up to 50,000 chars
• Voice input — global dictation hotkey, push-to-talk, auto-paste, in-app mic, Whisper STT
• Agent voice output — voicebox.speak MCP tool — Claude Code, Cursor, Cline speak in cloned voices
• Voice personalities — free-form personas + bundled local LLM for Compose/Rewrite/Respond (also MCP-exposed)
• API-first — REST API + built-in MCP server
• Runs everywhere — macOS MLX/Metal, Windows CUDA, Linux, AMD ROCm, Intel Arc, Docker

🎨 Why this matters for an agent workflow

Bridges the input and output halves of voice I/O. Existing cloud incumbents sit on opposite sides — Voicebox does both, glued by a local LLM. MCP integration means a Claude Code session can literally talk back to you in a voice you own — useful for Pulse-style background notifications or Personal AI Infrastructure DA personas.

Reasoning for

Local voice cloning for personal projects, dictation across all apps, giving any MCP-aware agent a configurable voice. The privacy story is the killer feature — ElevenLabs has every word you’ve cloned; Voicebox keeps it on the box.

Alternatives considered

• ElevenLabs — best quality, cloud-only, paid
• WisprFlow — STT-only, cloud
• Whisper.cpp + Piper — DIY local stack, no GUI, no MCP

🔗 Links

• Repo: https://github.com/jamiepine/voicebox
• Site: https://voicebox.sh/
• Docs: https://docs.voicebox.sh/

🔗 Related notes

• TTS Engines Comparison (Polish) — where Voicebox’s TTS engines (Kokoro, Chatterbox, Qwen3-TTS) sit vs cloud rivals for Polish
• Pocket TTS — Kyutai’s 100M CPU-only TTS engine; candidate 8th engine, same local-first philosophy
• Claude Code — primary MCP host
• Personal AI Infrastructure — DA voice personas
• Hermes Agent — agent that benefits from spoken output
• Paperclip — agent orchestration

---

Template: tool