Voicebox is a local-first desktop voice studio for macOS, Windows, and Linux that covers both directions of voice: it clones voices and generates speech, and it transcribes dictation into any app. It positions itself as a local alternative to ElevenLabs on the output side and WisprFlow on the input side, and models, voice data, and recordings stay on the machine.
Speech generation switches per request between eight TTS engines: Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, and Kokoro. Together they cover 23 languages, zero-shot cloning from a short reference sample, 50+ preset voices, voices built from a written description, and paralinguistic tags such as laughs and sighs (Chatterbox Turbo only). Long text is split at sentence boundaries and crossfaded, every generation keeps versioned takes with lineage, and a multi-track Stories editor arranges several voices into conversations or podcasts.
Dictation runs Whisper locally behind a global push-to-talk or toggle hotkey, pastes into the focused text field on macOS, and can pass the transcript through a bundled local Qwen3 LLM to remove filler and false starts. The same LLM powers voice personalities: a free-form persona attached to a voice profile composes in-character lines or rewrites input before it is spoken. Every recording lands in a Captures tab with its transcript, ready to re-transcribe, refine, or promote to a voice sample.
For agents, Voicebox exposes a REST API and a built-in MCP server with four tools (speak, transcribe, list captures, and list profiles). An MCP-aware agent such as Claude Code, Cursor, or Cline can speak task completions in a cloned voice, each client can be pinned to its own voice, and an on-screen pill shows whenever an agent is speaking. The desktop shell is Tauri with a React frontend and a FastAPI backend, and inference runs on MLX for Apple Silicon or PyTorch for CUDA, ROCm, DirectML, Intel Arc, or CPU.
Features
- Eight TTS engines: Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, and Kokoro, switchable per generation
- Voice cloning and presets: zero-shot cloning from a reference sample with multi-sample profiles, 50+ curated preset voices, and voices designed from a text description
- Expressive speech: paralinguistic tags with Chatterbox Turbo and natural-language delivery instructions with the Qwen engines
- Post-processing effects: pitch shift, reverb, delay, chorus, compression, gain, and filters, with built-in and custom presets per profile
- Long-form generation: sentence-aware chunking with crossfades for texts up to 50,000 characters, plus a non-blocking generation queue
- Generation versions: original, effects, and re-seeded takes with source lineage and favorites
- Stories editor: a multi-track timeline for conversations, podcasts, and narratives with trimming, splitting, and synchronized playback
- Global dictation: rebindable push-to-talk and toggle hotkeys, paste into the focused field on macOS, and optional local LLM cleanup
- Captures: every recording kept with its transcript, re-transcribable at any Whisper size and promotable to a voice sample
- Voice personalities: a persona per voice profile that composes or rewrites lines in character through a bundled local Qwen3 LLM
- Agent voice output: an MCP server and REST API that let Claude Code, Cursor, Cline, and other agents speak in a cloned voice, with per-agent voice binding
- GPU support: MLX on Apple Silicon and PyTorch on CUDA, ROCm, DirectML, Intel Arc, or CPU, with per-model unloading
