Voicebox

Voicebox is an open-source desktop app (macOS, Windows, Linux) that bundles voice cloning, multi-engine TTS, Whisper-powered dictation, and an MCP server into one local-first package. It clones a target voice from as little as three seconds of audio (upload, mic, or system capture), runs inference on Metal, CUDA, ROCm, Intel Arc, or DirectML, and exposes a voicebox.speak tool so Claude Code, Cursor, Cline, and other MCP-aware agents can talk back in voices you own.

It pitches itself directly as a free, local alternative to ElevenLabs and WisprFlow, and the feature surface backs that up: seven TTS engines, a Stories timeline editor for multi-voice narratives, an audio effects pipeline (pitch, reverb, delay, compression) with per-profile defaults, generations up to 50,000 characters with auto-chunking and crossfades, and a global push-to-talk dictation pill that drops cleaned transcripts into any app. A local LLM (Qwen) optionally scrubs ums and self-corrections without rephrasing.

Voicebox is free and open-source, with a GitHub repo and an associated $VOICEBOX donate/token angle. Beyond MCP it also exposes a POST /speak HTTP endpoint for ACP, A2A, shell scripts, and custom harnesses. The catch is that it's a desktop-only workflow that wants real GPU horsepower for the larger Whisper and TTS models, and the celebrity-voice presets it ships with raise the usual ethical and consent questions around cloning real people.