A new AI model learns to judge robotic-sounding speech the way a human would, without needing Gemini's bill every time it scores a clip.
Researchers built Live-ProsodyJudge, a system that scores AI-generated speech for qualities like emotion, tone, and energy, the details that make a synthetic voice sound alive instead of flat. It works by distilling judgments from Gemini into a smaller model called Qwen3-Omni, making it cheap enough to run constantly during training instead of only occasionally. But the team found the judge was cutting corners: it tended to let one overall impression of a clip dictate every individual score, collapsing several distinct quality dimensions into a single lazy verdict. Their fix, called Decoupled-Live-ProsodyJudge, removes that overall-preference shortcut and retrains the model to justify each dimension's score with its own separate reasoning.
Training expressive, real-time text-to-speech systems requires constant, fine-grained feedback on thousands of samples, and that kind of volume is exactly what makes a proprietary judge too expensive to use at scale. The distilled version closes that gap without much of a quality hit: in testing it agreed with human raters more often than a single call to Gemini did, and when asked to pick the best of eight candidate voice clips, its choice landed in humans' top three picks 85.29% of the time. That is a real efficiency trade worth noting for anyone building live voice assistants or streaming avatars, where prosody usually lags behind raw speed.
It is also a quiet admission that even a top-tier model's judgment is too costly to lean on directly, so the workaround is a cheaper model trained to imitate it, one rated clip at a time.