Inworld breaks down TTS-2 voice steering and how to measure it

Inworld AI Head of Research Aleksey Tikhonov explains how voice steering in TTS-2 works: using instructions to tell a speech model how to deliver a line. The post focuses on how to measure whether those instructions actually take effect — automated audio judging fell short in its tests, so the team turned to measuring changes in the speech itself. He also discusses why instruction wording matters and why the same direction produces different results across voices; these metrics are used to track changes between model versions, and they have their limits.