Voicebox: Local Voice I/O

The user runs Voicebox, a local-first voice studio. It is wired into this Hermes install in up to two ways:

1. Pipeline providers (this plugin): when tts.provider: voicebox / stt.provider: voicebox are set, your spoken replies and all incoming voice-message transcription already flow through Voicebox automatically. You don't need to do anything for these.
2. MCP tools (if the voicebox MCP server is connected): tools named mcp__voicebox__voicebox_speak, voicebox_transcribe, voicebox_list_profiles, voicebox_list_captures. These are for deliberate voice actions and recall.

When to speak aloud (MCP speak)

• The user explicitly asks ("read this to me", "say it in Morgan's voice").
• The user is away from the screen (they told you they're cooking, driving, etc.) and the reply is short and useful to hear.
• Do NOT speak unprompted in group chats, for long content (summarize first, then offer), or repeatedly after the user ignored/stopped audio.
• The profile argument accepts a profile name or id — list them with voicebox_list_profiles if unsure. Respect an explicitly requested voice; otherwise omit profile and Voicebox uses the binding the user set for this client.

Captures = the user's dictation memory

Voicebox's Captures tab stores everything the user has dictated, with the original audio preserved alongside each transcript. When the user refers to something they said or dictated earlier ("that idea I recorded this morning", "my voice note about the launch"), search Captures via voicebox_list_captures before asking them to repeat it.

Transcription behavior you should expect

• Whisper runs locally inside Voicebox. First use of a new model size triggers a background download — Voicebox answers with a "still downloading" error. Wait a minute and retry; tell the user what's happening rather than reporting a generic failure.
• If Voicebox is unreachable, the app isn't running. Say exactly that — "your Voicebox app doesn't seem to be open" — instead of a vague error.

Voice selection quick reference

• Voices = Voicebox profiles (cloned or preset). Engines (qwen, kokoro, chatterbox, …) are per-profile defaults; only override the engine when the user asks for a specific one.
• Long text: fine to send — Voicebox chunks and crossfades automatically.