Nova 2.5 Sonic turn-taking: does a voiceover need it?

No, a rendered voiceover does not need turn-taking control. AWS lists controllable turn-taking, mixed voice and text in one session, and asynchronous tool calling as Nova 2.5 Sonic features, and all three describe a live conversation with a person who interrupts. A voiceover has one speaker, a fixed script and no interruptions, so a text-to-speech job that returns a file is the matching tool.

Nova 2.5 Sonic facts are from the AWS announcement read on 2026-10-13. The Sume side is from the repo's TTS docs and catalog on main.

How do I keep the exact words?

For fixed copy, Sume TTS can bind a job to an accepted script: transcript_source names a script_revision_id and sentence_ids, and the job returns a transcript_receipt with a submitted-text SHA-256. The docs say the receipt proves the text, not the pronunciation, so listen to the take.

When would I still pick a speech-to-speech model?

When the product is a call, a kiosk or an in-app assistant. Then the AWS page, not a render pipeline, is where to evaluate it. When the product is a 15 second ad with a script, choose a TTS model and a voice, and let captions and the timeline do the rest.