Cleanup

Whisper produces a raw transcript: punctuation is rough, filler words are kept, casing can be off. Fluister runs the raw text through a local LLM ("the cleanup model") that:

• Fixes punctuation and capitalization.
• Removes filler ("uh", "um", "like", "you know").
• Keeps the meaning. Never adds or removes information.

The cleanup model is a bundled GGUF file (currently Llama 3.2 3B Instruct, Q4_K_M, 2.0 GB) run via a local llama-server (llama.cpp's HTTP server). It is not internet-connected.

Profiles

A profile is a named cleanup style. Each profile has a custom system prompt that nudges the cleanup model toward a particular tone.

Examples:

• Default — neutral, professional, conservative edits.
• Email — second-person, polite, paragraph breaks.
• Chat — casual, lowercase, no period at the end.
• Code comment — first-person, terse, fits in 80 columns.

Profiles live as markdown files in the vault. Open the vault folder (Settings → Storage → Open in Finder) and edit profiles/<name>.md in any text editor. Each file has a style_prompt field, a name, a description, and an optional list of app bindings (bundle identifiers like com.apple.mail). When you dictate into an app with a binding, that profile is used automatically.

Vocabulary

Whisper hears phonetically. It misspells unusual words ("OAuth" → "O-Off", "Vercel" → "her sell"). The vocabulary is a list of canonical terms plus the phonetic variants Whisper tends to produce instead.

Vocabulary entries live in the vault too, in vocabulary/terms.md. The canonical spelling is required. Aliases are optional (empty aliases just bias the model toward recognising the word at all). AI clients can also add entries via the add_vocabulary MCP tool.