What Voicebox does
Voicebox is a local-first AI voice app that pitches itself as a free, open-source alternative to ElevenLabs (speech output) and Wispr Flow (dictation) in one. You clone a voice from about 3 seconds of audio (an uploaded WAV, MP3, FLAC or WebM file, a mic recording of up to 30 seconds, or captured system audio), then generate speech of up to 50,000 characters per run; long text is split at sentence boundaries and crossfaded. You can switch engine per generation: the site lists Qwen3-TTS, Qwen CustomVoice, Chatterbox, Chatterbox Turbo, LuxTTS, TADA (Hume AI) and Kokoro, and the README adds Qwen VoiceDesign. Language coverage depends on the engine, from 10 languages up to 23 with Chatterbox. A Stories timeline editor handles multi-voice narratives, and an effects pipeline adds pitch shift, reverb, delay and compression.
For input, hold a global hotkey, speak and release: Whisper (99 languages) transcribes and, on macOS, pastes into the focused field, with optional clean-up of filler words by a local Qwen3 model. For agents, any MCP client (the site names Claude Code, Cursor and Cline) can call a voicebox.speak tool on 127.0.0.1:17493, and a local REST API exposes generation, profiles and history. Downloads are a DMG for Apple Silicon and Intel Macs and an MSI for Windows; Linux users build from source or use Docker. Acceleration covers Metal/MLX on Apple Silicon, CUDA, ROCm, Intel Arc and DirectML, or a remote machine.
