Controlling pronunciation
TTS engines read what is literally on the page, so you control pronunciation by controlling spelling. Write out numbers the way you want them spoken — 'nineteen ninety-five' reads very differently from the same digits. Force acronyms with capitals and hyphens: 'S-Q-L' when you want the letters, 'sequel' when you want the word.
For names and coined words the engine mangles, respell them phonetically — 'Siobhan' as 'shiv-AWN', a product name broken into stressed syllables — and keep a short pronunciation key at the top of your script file so every future generation uses the same fixes.
Pacing with punctuation
Punctuation is the pacing control surface. Commas insert short lifts, periods land full stops, ellipses and dashes read as hesitation or a held beat, and a paragraph break gives the voice a genuine reset. If a passage rushes, add punctuation; if it plods, remove some and join the sentences.
Structure matters at the document level too. Break long text into sections and generate each separately: a bad take only costs you one section, retakes stay cheap, and you can vary voices between sections — a different preset for chapter headings than for body text, for example.
What people voice with it
Course creators narrate lessons without recording sessions, so a curriculum update means regenerating one paragraph rather than re-recording a module. Video makers generate scratch VO to time their edits, then keep the TTS take when it holds up. Writers proof long drafts by ear, catching repetition and rhythm problems that silent reading misses.
The audio also feeds Nidhogg's video side directly: a TTS track can drive a talking avatar through InfiniTalk, or replace the dialogue in an existing clip through LatentSync — which turns a text edit into a video fix without a reshoot.