Voicebox setup: voice cloning, dictation and Claude Code

This tutorial covers Voicebox setup on an Apple Silicon Mac, from downloading the app and speech models to creating a voice profile. The presenter uses a 24-second recording with matching reference text and selects the recording's language before generating speech. The selected engines are Qwen3-TTS 1.7B and Chatterbox Turbo, with Whisper medium for transcription.

The walkthrough tests expressive speech tags for laughter and sighs, then applies a radio effect to generated audio. Chatterbox Turbo supports the expressive tags used here, though the presenter reports an odd accent in the result. A two-track podcast example shows how the Stories editor handles playback, reordering and regeneration.

For dictation, the presenter configures a hold-to-talk shortcut, automatic pasting and optional transcript cleanup through a language model. The Claude Code setup uses Voicebox as an MCP server. When the copied command returns an unknown URL option error, the demonstrated fix removes the URL flag while keeping the address.

The presenter describes Voicebox as free, open source and able to work offline on local hardware. The supplied metadata identifies an MIT license. Performance depends on the computer and model size. Both the first speech generation and first dictation attempt fail in this demonstration; restarting the app resolves each failure.