Skip to main content
Kalam Documentation

Models and Sense

Open Dictation → Models to choose the shared speech and language engines behind Plain dictation and every profile that inherits from it, and to configure Sense background checks.

Placeholder until dictation-models.webp is added under apps/docs/static/img/screenshots/.

Shared models and Plain dictation​

Plain dictation is the baseline profile. It uses the shared speech engine selected under Models, skips AI polish, and skips screen context so quick captures stay fast and predictable.

Other dictation profiles (Email, Notes, Coding, or custom profiles) can either inherit the shared engine choices from Models or override them per profile on Dictation profiles.

Setting on ModelsApplies toDetails
Speech enginePlain dictation and inheriting profilesPick On-device (SenseVoice Small, Whisper Base, Parakeet TDT 0.6B, or Moonshine Small) or Cloud (any connected provider).
AI polish modelProfiles with Improve enabledSelects the default language model used when a profile rewrites or formats speech.

Sense background intelligence​

The header bar of Dictation displays live status chips for Sense—three background safeguards that inspect each take automatically without extra hotkeys:

Sense featureWhat it does
Clean takeStrips filler words, false starts, and obvious microphone artifacts before formatting.
Polish checkGuards against accidental AI rewrites by verifying that the polished result stayed faithful to what you actually spoke; if the check flags drift, Kalam falls back to your clean raw transcript.
Smart profile routingWatches the active app and window context (unless blocked by Sensitive apps) so takes automatically follow the matching dictation profile.

Select any Sense chip in the Dictation header to jump directly to its card on the Models screen and view live before/after examples or toggle it on or off.