AI Voices Just Got Better at Not Forgetting Who They Are

Imagine paying for an AI narrator for your audiobook, and halfway through chapter three the voice quietly turns into a different person. That is a real problem in speech AI today, and a team of researchers from China say they have a fix.

Here is what is going on. Modern speech generators break sound into small chunks and store them as numbers. Most systems round those numbers into neat categories — fast and stable, but it throws away fine details like emotion, breath, and tiny pronunciation quirks. The alternative is to keep the numbers raw, which sounds richer but lets small errors pile up. Over a long script, those errors snowball, and the voice wanders off.

The team's trick is to squeeze that raw memory onto the surface of a shape called a hypersphere — think of it as forcing every note of a song onto the rim of a ball rather than anywhere inside it. Because the AI can only move around that surface, it can never drift too far from where it started. They call the method SphereVAE.

Tested inside a speech system called VoxCPM, it made noticeably fewer mistakes on the actual words spoken, kept the same speaker likeness, and held a consistent voice much longer. The honest catch: the audio is not a perfect replica of the original recording, because the AI has less room to store fine detail. For audiobooks, podcasts, dubbing, and voice assistants, that is usually a trade worth making — most listeners notice a voice changing far more than they notice a slightly softened vowel. The paper was accepted to NCMMSC 2026, a speech-processing conference.