Native-speaker review for an AI voiceover: a 10-minute checklist

Decagon validates locale voices with native speakers. A checklist to do the same review on any AI voiceover before you publish it in another language.

5 min readSume
All posts

Have a native speaker review every AI voiceover before you publish it in a new language, and give them a short checklist so the review takes about ten minutes. Decagon says Voice 3 ships locale-specific voices validated by native speakers, and the same step belongs in your own process.

What to ask the reviewer

Send the audio, the script and the target locale together, then ask for answers to these:

  • Does the accent match the locale, for example Mexican or European Spanish, Brazilian or European Portuguese?
  • Are numbers, dates, prices and phone numbers read in the form a local would say?
  • Are brand names and product terms pronounced acceptably?
  • Does the pace and tone suit the audience?
  • Does anything sound like another language's rules applied to this one?

Pick locales deliberately

Microsoft's docs list MAI-Voice-2.1 voices by locale, such as es-ES and es-MX, or pt-BR and pt-PT, so the region is part of the voice name. Treat the locale as a casting decision and write it into the brief.

Locale choices that change the read (read 2026-10-07)
LanguageLocale pairWhy it matters
Spanishes-ES, es-MXDifferent vocabulary and rhythm
Portuguesept-BR, pt-PTDifferent pronunciation and idiom
Englishen-US, en-GB, en-AU, en-INSpelling, units, idioms
Chinesezh-CNSimplified script voices

Fix loop on Sume

If the reviewer flags one sentence, retake only that sentence: generate it as its own TTS job (one tts_create per sentence is the pattern the MCP docs describe) and rejoin the takes with timeline audio concat for $0.01 flat per job, instead of regenerating the whole script.

Before the retake, check the language field. On hosted MCP, a known voice and language mismatch (409 tts_voice_language_mismatch) comes back as a warning that needs confirmation instead of a submit, so the wrong-language error is caught before you pay.

After the audio is approved, burn captions with the same approved script_text so the captions match what the reviewer heard. A caption job reserves $0.20 for videos up to 60 seconds.

Record the sign-off

Keep the approved audio URL, the script text and the reviewer's name in your own project record. A short sign-off line prevents the next team from regenerating an approved file by accident.

Sources

Related posts

More in Use cases

All Use cases posts

Written by Sume