Write for the ear, not the eye
Text written to be read silently sounds stiff when spoken. Before generating, convert it to spoken English: shorter sentences, contractions, one idea per breath. A reliable test is reading the script aloud yourself — anywhere you stumble or run out of air, the voice will too. If a sentence needs a second pass to parse on paper, split it.
Punctuation is your direction to the voice. Commas produce short lifts, periods land full stops, and paragraph breaks give longer resets — so an ad script punched into short fragments reads with energy, while long flowing clauses read calm and documentary-like. Spell out anything ambiguous: 'twenty twenty-six' instead of the digits, 'S-E-O' when you want letters, and a phonetic respelling for names the engine trips on.
Choosing a voice and an engine
Match a voice's baseline character to the job before trying to force a read: warm and conversational for explainers and podcasts, crisp and forward for ads and promos, measured and neutral for corporate narration. Each of the five engines carries a different roster, so audition your actual opening paragraph — not a throwaway test sentence — on a handful of voices and keep the one whose default rhythm already fits.
Generations cost 2–4 credits each, which changes how you work: auditioning five voices is a trivial spend, and there is no session fee or minimum. Once a voice works, note the engine and voice name and reuse the exact pair for every future script, so a series sounds like one narrator across episodes.
From voiceover to talking video
Generated audio plugs straight into Nidhogg's lipsync tools. InfiniTalk animates a still portrait — a founder photo, a brand mascot, a generated character — into a talking head that speaks your voiceover. LatentSync and Sync Lipsync go the other way: give them an existing video plus the new audio, and they re-sync the speaker's lips to match the replacement track.
For voice-under-picture work, time the script before you generate visuals. Spoken narration runs around 150 words per minute, and Nidhogg video generations are short clips of a few seconds each — so a 40-word paragraph pairs naturally with two or three clips, or with one multi-shot generation on Seedance 2 or Kling 3.
Retakes without a booth
The practical win over recorded voiceover is not the first take — it is take fifteen. A mispronounced product name, a script change from legal, a new price point: regenerate the single affected line for a couple of credits instead of re-booking a session, then splice it into your edit.
Keep long scripts split into labeled sections and generate each separately. Section-level files make retakes surgical, keep any one generation short and clean, and let you re-voice one chapter without touching the rest.