Consumer tech delivers 500K voice prompts across 8 locales

Overview

Voice assistants don't ship in new languages without expressive, native-recorded prompts behind them.

A Global Consumer Tech Company needed to improve its virtual assistant and introduce new language capabilities ” and that meant collecting sensitive and non-sensitive short prompts recorded in expressive native voice across 8 locales.

Firstsource recruited 5,000+ native speakers across 8 countries and delivered 500,000+ prompts recorded, transcribed, and annotated end-to-end.

This was Intelligence that Operates: a multilingual speech pipeline that stood up three new processes ” recording, transcription, and annotation ” and accelerated model development by half.

Challenges

• Crowd-sourced voice doesn't sound like the people the model serves. A virtual assistant trained on stilted recordings will sound stilted back. Expressive prompts have to come from native speakers in-locale, not voice actors reading scripts.
• 8 locales mean 8 different recruiting, briefing, and quality challenges. UK English isn't US English; Mexican Spanish isn't Spain Spanish; Canadian French isn't France French. Each locale needs its own briefing and review discipline.
• Sensitive content can't be handed to general annotation pools. A meaningful share of prompts involves sensitive material ” recording, transcription, and annotation all have to run on workflows that protect both the speakers and the data.

How We Made It Happen

We stood up the recording, transcription, and annotation pipeline from scratch as one program across all 8 locales.

• Native speakers in-country across 8 locales. 5,000+ speakers recruited across UK, US, Canada, France, Germany, Spain, Mexico, and Italy ” briefed for expressive delivery rather than read-aloud.
• OTS Data delivered as a fully labeled corpus. Recording, transcription, and annotation ran together, not as three handoffs ” 500,000+ prompts arrived production-ready.
• One quality discipline across the program. 95%+ first-time-right transcription and 97%+ recording quality held across all 8 locales on the same review framework.

Conclusion

A voice assistant only sounds natural in a new language if the data behind it was. Firstsource ran the recording, transcription, and annotation pipeline at multilingual scale ” turning speech data collection into Intelligence that Operates.