Voicebox

Voicebox is a local-first voice studio that covers both halves of the voice I/O loop on one machine: text-to-speech output (voice cloning and generation) and speech-to-text input (dictation into any application). Models, voice data, and captures stay on the machine; no cloud service is involved.

You clone a voice from a few seconds of reference audio or pick from preset voices, then generate speech in up to 23 languages across seven switchable TTS engines. Long scripts are split at sentence boundaries, generated per chunk, and crossfaded, and every generation keeps versions (original, effects chains, re-seeded takes) with lineage tracking. A Stories editor composes multi-voice timelines for conversations and podcasts.

On the input side, a global hotkey records speech anywhere on the system and, on macOS, pastes the Whisper transcript into the focused text field; a bundled local LLM can clean up fillers before paste. For agents, a REST API and a built-in MCP server expose the same capabilities, so an MCP-aware agent such as Claude Code, Cursor, or Cline can speak through a cloned voice with a single voicebox.speak tool call, and voice profiles can carry a persona used to compose, rewrite, or respond to text.

Engine	Languages	Notes
Qwen3-TTS (0.6B / 1.7B)	10	multilingual cloning with delivery instructions
Qwen CustomVoice	10	9 preset voices, natural-language delivery control, no reference audio
LuxTTS	English	~1 GB VRAM, 48 kHz output
Chatterbox Multilingual	23	widest language coverage
Chatterbox Turbo	English	350M model with paralinguistic tags such as [laugh] and [sigh]
TADA (1B / 3B)	10	HumeAI speech-language model for long coherent audio
Kokoro	8	50 preset voices, 82M model, CPU inference

Features

• Multi-engine cloning: seven TTS engines switchable per generation, zero-shot cloning from a reference sample or 50+ preset voices
• Paralinguistic tags: inline [laugh], [gasp], [sigh], and similar tags interpreted by Chatterbox Turbo
• Post-processing effects: pitch shift, reverb, delay, chorus, compressor, gain, and high/low-pass filters, with built-in and custom presets assignable per profile
• Unlimited length: configurable auto-chunking (100–5,000 chars) with crossfade, up to 50,000 characters of input
• Generation versions: originals, effects versions, re-seeded takes, favourites, and source tracking
• Async queue: non-blocking serial generation with SSE status streaming, retries, and crash recovery
• Voice profiles: create from files or in-app recording, multi-sample cloning, import/export, per-profile default effects
• Stories editor: multi-track timeline with drag-and-drop, trimming, splitting, and version pinning per clip
• Global dictation: rebindable push-to-talk and toggle chords, target-aware paste on macOS with clipboard restore, on-screen status pill
• Speech-to-text: Whisper sizes from Base to Large plus Turbo, on MLX or PyTorch (CUDA, ROCm, DirectML, CPU)
• Captures: every dictation and recording kept with its transcript, re-transcribable, editable, and promotable to a voice sample
• Agent integration: REST API and MCP server; voicebox.speak gives any MCP client a voice, and persona modes are callable over MCP