Reproduce the same AI voiceover later: model id, voice, settings

To recreate a narration line later you need five things: the model id, the voice, the language, the output format and any settings you sent. Sume records all of them on a completed text-to-speech job, so the routine is to read them off the job and store them with the audio, not to rely on memory or a prompt file.

Why a model name is not enough

Vendors update models in place. Cartesia's documentation lists sonic-3.6 as a rolling id, sonic-3.6-2026-08-27 as a frozen dated snapshot, and sonic-preview as a preview id. A rolling id can change what it sounds like when the vendor ships an update, while a dated snapshot stays fixed. If a client asks for the same read of a line next spring, only the dated form gives you a fair chance of matching it.

What Sume stores on the job

Per the Jobs and results page, a completed text_to_speech job records how its audio was made: model_id for the engine, voice as { "mode": "id", "id": "…" }, language, output_format, and the settings it was synthesized with, generation_config and speed. Each setting is null when the request did not send it. The docs say to read these from the job to make the next line sound the same.

The fields tell you what was requested. They are the parameters of the call, not a guarantee that two runs will be bit-identical, so keep the original audio file as the master and treat regeneration as a fallback.

curl https://api.sume.com/v1/jobs/job_123 \
  -H "Authorization: Bearer $SUME_API_KEY"

curl https://api.sume.com/v1/jobs/job_123/result \
  -H "Authorization: Bearer $SUME_API_KEY"

A routine that takes a minute per project

When a voiceover is approved, do these four things once.

• Copy the job id and the model, voice, language, format, config and speed fields into your project notes.
• Download the finished audio from its media.sume.com URL and keep it as the master.
• Note whether the model id was a rolling or a dated form.
• Write down the script text exactly as sent, including any pronunciation edits.

When the next line must match

For a follow-up line in the same project, send the same voice and settings, and the same model id unless you intend to change it. Where the voice is part of the brand, keep it constant across lines and change only the text. Compare the new clip against the master by ear before you cut it into a timeline.