voicebox

Voicebox is an open-source, local-first AI voice studio positioned as an all-in-one alternative to ElevenLabs and WisprFlow. It covers both directions of voice input and output: on one hand it provides voice cloning and voice generation, and on the other it provides dictation and transcription, and uses a built-in local large model to perform text polishing and personalized rewriting. The official emphasis is on complete privacy: models, voice data, and captured content never leave the local machine.

In terms of use cases, Voicebox can be used for voice cloning, speech synthesis, dictation input, transcription archiving, agent voice output, and multi-track audio production for podcasts, dialogue, and narrative content. It is aimed both at individuals for everyday dictation and at developers who need to give AI agents a voice.

Main capabilities include:

• 8 TTS engines: Qwen3-TTS, Qwen CustomVoice, Qwen VoiceDesign, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, Kokoro.
• Voice cloning and preset voices: supports zero-shot cloning based on reference samples, as well as more than 50 curated preset voices provided through Kokoro and Qwen CustomVoice.
• Support for 23 languages, covering English, Arabic, Japanese, Hindi, Swahili, and more.
• Post-processing effects: 8 audio effects provided by the Spotify pedalboard library, including pitch shifting, reverb, delay, chorus, compression, and filtering; built-in Robotic, Radio, Echo Chamber, and Deep Voice presets, support for custom presets, and setting a default effect chain per voice profile.
• Expressive speech: Chatterbox Turbo supports paralinguistic tags such as [laugh], [sigh], and [gasp] (some other engines will read them out as literal text); Qwen CustomVoice supports controlling expression in natural language; Qwen VoiceDesign supports describing a voice in plain language.
• Unlimited-length generation: automatically chunks by sentence boundaries, generates independently, and crossfades and stitches them together, with support for a chunk limit of 100–5,000 characters and crossfade adjustment of 0–200ms, a maximum text length of 50,000 characters, and intelligent segmentation that can recognize abbreviations, CJK punctuation, and [tags].
• Generation version management: each generation retains the original version, effFruit versions, takes from different seeds, and records the source lineage, with favorites supported.
• Asynchronous generation queue: non-blocking submission, serial execution to avoid GPU contention, real-time SSE status stream, retry on failure, and automatic recovery of stale tasks after a crash on startup.
• Voice profile management: create from audio files or in-app recording, import/export, multi-sample support, set a default effect chain per profile, and support descriptions and language tags.
• Story editor: multi-track timeline with drag-and-drop, inline trimming and splitting, synchronized playhead, and per-clip pinned versions, for dialogue, podcasts, and narrative production.
• Global dictation and voice input: configurable chord key bindings, support for hold-to-talk and click-to-toggle; on macOS, tapping Space while holding upgrades seamlessly to toggle mode; macOS supports target-aware paste (verifies and injects into the focused text box via Accessibility, and atomically saves/restores the clipboard); on first run, guides the user to grant macOS Accessibility and Input Monitoring permissions; every Voicebox text box has an in-app microphone button; optional LLM polishing removes ums, stutters, and false starts.
• Transcription and Capture management: based on OpenAI Whisper, running on MLX (Apple Silicon) or PyTorch (CUDA/ROCM/DirectML/CPU). Every dictation, in-app recording, and uploaded audio goes into the Captures tab, and the original audio and transcription are always retained; you can replay, re-transcribe with any Whisper size, re-polish with different flags, edit inline, read aloud with a cloned voice in one click, or promote the audio and transcription to a voice reference sample. Locally captured content is stored in the Voicebox data directory, with folder shortcuts provided in settings.
• Agent voice output: provides an MCP server, so a single tool call (voicebox_speak) lets MCP-enabled agents speak with a cloned voice; it also exposes POST /speak for non-MCP scenarios such as ACP, A2A, and shell scripts. Supports binding voices per agent (e.g., Claude Code bound to Morgan, Cursor bound to Scarlett), and uses the last_seen_at timestamp to confirm that installation is in effect; all agent-initiated speech displays an on-screen floating pill and will not be read aloud silently in the background. Supports both HTTP and stdio transports.
• Voice personality: any voice profile can be given its owntext personality, driven by a local Qwen3 LLM, providing two operations: Compose (generate lines that fit the character) and Speak in character (rewrite in character and then TTS); agents can use the same rewrite path by passing personality: true to voicebox_speak. The local LLM can be Qwen3 0.6B / 1.7B / 4B, sharing the runtime with TTS.
• Model management: unload per model to free GPU memory without deleting downloads, support a custom model directory via VOICEBOX_MODELS_DIR, model folder migration with progress tracking, and a download cancel/clear UI.

In terms of usage: it can be obtained through the official website, documentation, and download page; there is currently no prebuilt binary for Linux, so it must be built from source according to voicebox.sh/linux-install. The project provides a REST API, and the full API documentation is located at http://127.0.0.1:17493/docs。MCP Installation supports a one-line command for Claude Code and configuration for any HTTP MCP client, with a stdio fallback that can point to the built-in voicebox-mcp binary. Four MCP tools are published: voicebox_speak, voicebox_transcribe, voicebox_list_captures, voicebox_list_profiles; per-client voice binding is managed in Voicebox → Settings → MCP. Local builds require installing Bun, Rust, Python 3.12, Tauri prerequisites, and Xcode on macOS, and use just commands (brew install just or cargo install just; run just --list to view all commands); the repository root includes a preconfigured .mcp.json, and running Claude Code inside the checked-out directory will automatically pick up the Voicebox MCP tools. There is a step-by-step guide for adding new voice models, and it is optimized for AI coding agents.

Use cases include: agent development loops (asking questions by voice, hearing answers in a cloned voice), interactive characters for games and narrative tools, podcast production, conversational and narrative audio, accessibility aids, voice assistants, and content automation.

Regarding limitations: the project is under the MIT license, free and open source, with no splash screens, banners, or interruptions in the app. Precompiled Linux binaries are not yet available. Paralinguistic tags are supported only by Chatterbox Turbo; some other engines will read them out literally. More engines (Parakeet v3, Qwen3-ASR) are planned. For project engineering status, open issues, and the priority queue, see docs/PROJECT_STATUS.md; for contribution guidelines, see CONTRIBUTING.md; for security reports, see SECURITY.md.