SIGMAAI
Case studies
Case study

Helping AI Assistants Sound More Human

Persona and dialect review of a real-time voice assistant across eighteen locales, ahead of studio recording.

Helping AI Assistants Sound More Human
DELIVERYPRICINGCOVERAGEKPIS
Remote, native-local networkPer hour18 locales, EN-US referenceCultural naturalness, persona consistency, studio go/no-go

The Problem

A real-time voice assistant ships first in US English, then gets adapted for every market it launches in next.

  • Translating the script gets the words right. It doesn't guarantee the assistant still sounds like someone from that market, rather than an American voice speaking their language.
  • Sigma's linguistic consultants had a name for what happens when it goes wrong: translated American English.
  • Technically accurate, culturally off, and expensive to catch once a studio has already recorded it.

The Insight

Nobody can write a rubric for "sounds native." A consultant either recognizes the voice as belonging to their market, or they don't.

  • The review started the way any native listener actually tests a voice: consultants opened a live conversation with the assistant first, before looking at the US reference script, to hear its tone and pacing on its own terms.
  • Only then did they compare it against the English original, line by line, and answer one question per line: would this work for a recording studio in your market today, or does it need a different approach?
  • No rubric meant no false precision. A consultant who said no still had to explain why and offer an alternative in their own words, not check a box.

The Stakes

  • Every script that reads correctly on paper but sounds foreign once spoken passes text QA and lands on the studio floor, where the fix is a re-record, not a re-edit.
  • A voice that sounds translated in one market carries the same trust cost as a mistranslated word, except nothing in an automated pipeline catches it. Every word technically checks out.

What We Built

  • Sourced native-market linguistic consultants across 18 locales (bn-IN, de-DE, fr-FR, id-ID, it-IT, pl-PL, pt-BR, ru-RU, th-TH, vi-VN, cmn-CN, mr-IN, nl-NL, ro-RO, ta-IN, te-IN, tr-TR, uk-UA), screened for high-level language proficiency and depth of cultural fluency in their home market, not bilingual ability alone.
  • Ran the review as a live conversation first, script comparison second, so judgment was anchored in how the assistant actually behaved, not just in a transcript.
  • Delivered a go/no-go call per line, with native-language rewrite alternatives wherever the US persona didn't translate.
  • Built the workflow without a formal guideline document. The brief moved through a kickoff conversation instead of a written spec, and Sigma still delivered a review other locales could repeat.
  • Scaled the same model to additional locales once the first wave proved out, using the same familiarize-compare-report structure each time.

Why it Matters Now

  • For speech and voice leads: a working test for real-time conversational flow, not just static translation quality.
  • For localization and audio production leads: a native go/no-go call before studio dollars get spent, not after.
  • For model alignment and safety leads: a way to catch cultural bugs and persona inconsistency before they reach a consumer market, without waiting for user complaints to surface them.

A translation can be perfect and still sound like it's from somewhere else.