SEB — Voice & Butler

SEB (working name; the wake word "Sebastian" and its backronym are user-editable) is SEBcentral’s butler. It comes in two parts: SEB Home, the opening overview page (orb plus panels, a visual command centre), and SEB Assistant, voice and dictation that follows you across the whole app via the ever-present bottom toolbar.

Three layers: what is software vs what is AI

• Voice layer — software (free, offline, cross-platform). Speech-to-text uses whisper.cpp locally; text-to-speech uses kokoro locally (default voice British male "George"). Both run on Mac and Windows, free and unlimited.
• Command engine — software (no model at all). A regex intent router plus skills runs all task commands and dictation — instant, reliable and free. This is most of SEB.
• Reasoning AI — the LLM (integrated). Only for freeform, chatty requests, using your connected cloud AI (Claude via Settings → Assistant). There is no silent local fallback: with no key, freeform says "connect your AI" while commands still work fully. The local Ollama path only runs if you explicitly pick engine = Local.

Principle: voice is not AI. Voice and commands are the always-on free core; the LLM is the integrated brain for the open-ended 10%.

Opt-in install

Sebastian (voice) is an opt-in install, not bundled. The base app is the CRM plus the normal star AI chat assistant. The Sebastian tab holds an installer until set up; declining loses nothing — the tab just waits. The wizard covers what he is, downloading the voice models (~200 MB, with a progress bar), choosing his brain (cloud Claude key or a guided Ollama install), picking a voice and wake word, and a test.

Reach and dictation

• Dictation (free, whisper) — insert speech-to-text at the cursor, anywhere in the app.
• Wake / listen — the SEB icon (manual) or the wake word "Sebastian" (opt-in, always-on).
• Speech-to-text and text-to-speech are free, local and unlimited; only generative requests use a model.