At a Glance:
Voicebox is a local-first ElevenLabs and WisprFlow alternative combining seven-engine voice cloning, Whisper transcription, system-wide dictation, a multi-track stories editor, and REST or MCP access for agent voice output.
Overview:
Voicebox is an open-source AI voice studio that runs speech generation, transcription, dictation, and agent voice output locally. Users can clone voices from reference audio or select bundled presets, switch among seven TTS engines, process long text through automatic chunking, and arrange multiple voices on a timeline. Its global hotkey records speech for Whisper transcription, with direct paste into the focused field documented for macOS. A REST API and built-in MCP server let applications and compatible agents speak, transcribe, and browse stored captures or profiles.
Key Decision Points:
Platform packaging: Downloads are available for Apple Silicon and Intel macOS plus Windows, Docker is supported, and Linux currently requires a source build.
Local compute options: Inference uses MLX on Apple Silicon or PyTorch across CUDA, ROCm, DirectML, Intel XPU, and CPU, with model-specific memory and performance characteristics.
Voice input scope: The global dictation workflow supports configurable push-to-talk and toggle bindings, but target-aware automatic paste is specifically documented for macOS.
Engine tradeoffs: Seven TTS engines cover different languages and capabilities; only Chatterbox Turbo interprets the documented paralinguistic tags, while other engines may read those tags literally.
Developer interfaces: REST endpoints and HTTP or stdio MCP transports expose speech, transcription, captures, profiles, per-client voice bindings, and optional personality rewriting.
Core Features:
Seven-engine voice cloning: Switch between Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox variants, TADA, and Kokoro for cloning, preset voices, delivery control, or language coverage.
Long-form generation and versions: Split text at sentence boundaries, crossfade independently generated chunks, preserve originals, create takes, apply effects versions, and track their lineage.
Stories editor: Arrange conversations, podcasts, and narratives in a drag-and-drop multi-track timeline with trimming, splitting, synchronized playback, and pinned clip versions.
Dictation and captures: Record from a global hotkey or in-app microphone, transcribe with Whisper, refine transcripts locally, and retain paired audio and text in the Captures tab.
Audio effects: Preview and save chains using pitch shift, reverb, delay, chorus or flanger, compression, gain, and frequency filters.
Agent voice output: Let MCP-aware agents or REST clients speak through cloned profiles, assign voices per client, and surface playback through the desktop status pill.
Use Cases:
Clone a voice and generate multilingual speech with an engine chosen for the required language and delivery controls.
Dictate into applications on macOS, then replay, edit, re-transcribe, or refine the stored capture.
Assemble multi-speaker podcast, conversation, or narrative audio in the stories timeline.
Add cloned-voice speech and transcription to an application or an MCP-aware agent through local REST, HTTP MCP, or stdio MCP interfaces.
Open-Source Alternative Value:
Voicebox is an MIT-licensed alternative to ElevenLabs and WisprFlow that combines voice output and input in a locally running desktop application. The repository includes its Tauri interface, FastAPI backend, seven TTS integrations, Whisper transcription, SQLite storage, REST API, and HTTP or stdio MCP server. Developers can run it from source, use Docker, add new TTS engine adapters through the documented architecture, or connect applications and agents to the same local voice profiles and captures.
