voicebox: Create and Use AI Voices Locally

Overview

Voicebox brings speech generation and voice input together in a local-first desktop app. Users can create speech from text with different voice models, clone voices from audio samples, and dictate into other apps. It also connects these capabilities to AI agents through an API and MCP server.

It is a fit for people who want an integrated alternative to separate voice-generation and dictation tools, especially when keeping audio, voice profiles, and model processing on their own machine matters. Its broad feature set is useful for both hands-on audio creation and developer workflows.

Key Features

• Generate speech using seven TTS engines, including options for voice cloning, preset voices, and multiple languages.
• Dictate through a global hotkey or in-app microphone, with Whisper-based transcription and optional local LLM refinement.
• Manage voice profiles, recordings, generation versions, and multi-track story projects.
• Add audio effects and create longer outputs through automatic text chunking and crossfades.
• Integrate speech generation, transcription, profiles, and captures with a REST API and built-in MCP server.
• Run inference locally with platform-specific backends, including MLX on Apple Silicon and PyTorch options for supported GPUs and CPU.

Use Cases

• Content creators can produce narration, dialogue, and podcast material using cloned or preset voices.
• People who dictate frequently can use the global hotkey to enter speech into other desktop apps, particularly on macOS where automatic paste is supported.
• Developers can give MCP-aware agents a voice or add speech and transcription to their own tools through the local API.
• Accessibility and voice-assistance projects can combine dictation, transcription, and speech output with locally managed voice profiles.

Getting Started

For desktop downloads, visit the latest releases. For source setup and prerequisites, see the README.

Considerations

• Local inference keeps model processing on the user's machine, but performance and available engines depend on the platform, backend, and hardware. CPU operation is supported but may be slower.
• Pre-built downloads are listed for macOS and Windows, while the README says Linux binaries are not yet available and gives build-from-source instructions.
• Global dictation auto-paste is described as supported on macOS; Windows and Linux paste support is listed on the roadmap.
• The project combines a desktop app, Python backend, and multiple model runtimes, so development from source requires Bun, Rust, Python 3.11 or newer, and Tauri prerequisites.
• Voice cloning and agent speech involve impersonation risks. Use voice samples and cloned voices with appropriate permission.