voice & models

• wake & listen — an offline wake-word path runs even before the heavier models spin up, so “hey” works from a cold start.
• transcribe — dictation runs a local whisper model. base.en is the default; heavier sizes like medium.en are a switch away when you want more accuracy and have the gpu for it.
• think — the assistant’s language model runs locally through ollama. no request, token, or transcript ever leaves the machine.
• speak — replies are voiced locally, and the voice itself is selectable: ask for “the voice selector” and take your pick.

the assistant operates the machine, not just the conversation. the honest inventory, straight from its own capability list:

• windows — open and close the browser, notes, and the surveillance grid; open tabs, navigate to any site, run searches.
• appearance — change or randomize the background, its opacity and material; swap the orb; sync accent colors across your apps; undo any of it.
• voice — change its own voice via the selector.
• system — report hardware telemetry, resize and arrange windows, purge memory, clear the transcript, run macros, and remember anything you tell it to remember.
• live data — weather anywhere, breaking news, sports scores, stock and crypto prices, and wikipedia on any subject.

no particular phrasing is required — say it plainly and it happens. answers that carry data arrive as widgets.

• on smaller gpus, keep whisper at base.en — it’s fast and accurate enough for command speech.
• the language model is small and quantized by design: an assistant that answers in a beat beats a genius that answers in a minute.
• if the machine is busy, the orb tells you — it has distinct idle, listening, and speaking poses, so you always know whether it heard you.