AI Voices Invent Words — New Fix Cuts Errors Sevenfold

AI voices are everywhere — reading your audiobook, announcing your GPS turns, answering the phone for your bank. These systems work by predicting sound, piece by piece, from written text. Sometimes they get it wrong and invent words that were never in the script, a problem researchers call "hallucination." You hear it as a random word, a repeated syllable, or a sentence that dissolves into gibberish.

A team of researchers studied why this happens. Their surprising finding: it isn't simply a matter of the AI paying more attention to the text. When the link between text and speech gets too sharp, the system becomes brittle — much like over-tightening a guitar string makes it snap instead of sing. A moderate, well-calibrated connection produced the fewest errors. Too little attention was bad, but too much was just as bad.

Their fix is called AlignDPO. It's a training technique, not a new app you'd download. During training, the system is shown examples of correct output it should prefer, and the team added a small extra nudge that keeps the text-to-speech link at that sweet spot. Nothing changes when the AI actually runs, so it stays as fast and cheap to operate as before.

On a standard English test set, the method cut the rate of serious content hallucination from 4.4% to roughly 0.6% — about seven times fewer mistakes. Human listeners also rated it more natural than the earlier version. The catch: this is a research result on one English benchmark, presented at an academic speech conference, so it will take time to reach consumer apps. Still, it points toward AI voices you can trust for contracts, medical instructions, and news read aloud.