Frontier AI lab lands 300K multilingual voice evals

Overview

A voice model is only as good as it sounds outside English.

For a Frontier AI Lab expanding into priority post-English markets, scaling voice AI meant collecting native-expert preference signals across languages where intonation, prosody, and cultural register all matter.

Firstsource ran a multilingual voice preference program; 30,000 evaluations per language across 9+ languages, generating preference signals for model refinement at the scale needed to accelerate market launches.

This was Intelligence that Operates: native-expert preference signals feeding model iteration on a global multilingual cadence.

Challenges

• Post-English expansion needs more than translation. A voice model trained mostly on English data will sound flat, off-tempo, or culturally misaligned in other languages. Preference signals have to come from native speakers in-market, not from translators.
• 30,000 evaluations per language is a workforce problem, not a tooling problem. Sustaining preference data quality across 9+ languages requires culturally fluent evaluators who can hear the difference between two acceptable responses and pick the more natural one.
• Preference signals only matter if they reach training fast. A backlog of multilingual evaluations sitting in a QA queue stalls model iteration. The lab needed signals flowing on a cadence that matched its model release schedule.

How We Made It Happen

We built one global preference pipeline rather than 10 disconnected language programs.

• Native-expert evaluators in each market. Speakers were sourced and vetted in-country, briefed on the model's intent, and calibrated to a common quality bar before production work began.
• Expert Preference (RLHF) at multilingual scale. Preference signals on speech naturalness, controllability, and cultural alignment fed model refinement directly ” not a batch report at the end of the program.
• One pipeline, every language. A single program orchestrated across markets, with regional golden datasets curated by in-market experts.

Conclusion

Voice AI that ships globally has to sound right in every language, not just one. Firstsource ran multilingual voice preference end-to-end, turning model expansion into Intelligence that Operates.