New AI Trick Teaches Voice Assistants to Say Your Name Right

You've heard it: a GPS mangling a street name, a smart speaker stumbling over a surname, an audiobook narrator — an AI one, anyway — inventing a pronunciation for a medical term. That happens because most text-to-speech (software that turns written words into spoken audio) has no dictionary. When it meets a word it hasn't seen, it simply guesses. Fixing this usually means feeding the system hours of recorded speech or correcting words one at a time.

Shuhei Kato's new paper, submitted to a major audio conference, does neither. The idea is elegant: ask the AI to read a sentence containing a common word it already pronounces correctly. Then take that audio — the AI's own output — and reuse it as the teaching example for the same sentence, with that one word swapped out for a tagged, correctly-accented version. The AI becomes its own teacher. No new recordings, no human coaching.

The results are concrete. Human listeners, rating carefully, heard the intended accent on 89% of unfamiliar words, compared with 57% for an older Japanese spelling-based approach that can't express accent at all. Naturalness — whether the voice still sounds human — wasn't measurably damaged. The method also transferred to several other AI voice systems without retraining, improving pronunciation accuracy on 319 test words by 25 to 47 points over doing nothing.

The catch: it isn't universal. The trick worked well on one system, CosyVoice 2, getting accents right on two words out of three, but failed entirely on another called Irodori, and the paper explains why. The work is also focused on Japanese, so benefits for English names and place names are promising but not yet proven. Still, the direction is clear: AI voices that learn from themselves, and fewer awkward corrections in your audiobooks and customer-service calls.