Murmurtale Dialogue intelligence
In Fiction mode, a small on-device model reads each sentence and labels it as narration, spoken dialogue, or interior thought. Dialogue gets quotation marks; narration stays unquoted; direct interior thought is set in italics with its attribution tag left plain.
This model labels your sentences — it never rewrites them. A final pass balances quotation marks and repairs doubled punctuation so nothing comes through mangled.
You control it in Preferences → Dialogue intelligence: Off, On-device, or Cloud (Pro). The on-device engine and its model (about 2.5 GB) download on first use, with progress shown.
Interior thoughts
When dialogue intelligence detects direct interior thought, it formats it in italics rather than quotes. In rich-text apps the italics render directly; in plain-text fields Murmurtale falls back to the style you choose at Preferences → Character thoughts (Italics, Asterisks, Quotation marks, or Plain).
Paragraph breaks
Murmurtale listens for shifts that suggest a new speaker or beat and starts a new paragraph. This speaker-change detection is a heuristic: it inserts paragraph breaks only and never adds quotation marks on its own — that job belongs to dialogue intelligence.
You choose the spacing at Preferences → Paragraph style: a single return between paragraphs (the default, matching modern editors) or a blank line between them (classic manuscript spacing).
Honest limitationOn-device dialogue detection occasionally under-formats — a spoken line that continues into narration in one breath, or a short unattributed reply, may not get quoted. It errs toward leaving text alone rather than adding wrong quotes. The Cloud (Pro) engine handles these harder cases. See Known limitations.