Fixing Russian stress errors in text-to-speech for a language-learning site

A developer running a reading site for language learners found that the text-to-speech audio for Russian kept putting stress on the wrong syllables, which matters because stress can change a word's meaning. The cause was a cleaning step that stripped the stress marks before the text reached the voice, along with the common practice of writing е in place of ё. Restoring the marks and ё, and adding instructions about them, raised accuracy on the test set only modestly, but restoring ё brought its error rate down sharply. The developer built a pipeline that determines each word's stress from overrides, a learner glossary, and an open-source accent model, with an LLM reviewing disagreements, and re-generates clips that fail a phone-level stress check. The approach reduced wrong ё from about 40% to about 2% and other stress errors from 13% to under 1%, for roughly $35 of API time. Single-word clips that still fail after three attempts are sent to a human listener.