TTS evals are broken because the scores the field trusts do not match what people actually prefer in real conversations.

I think this is a solid critique because TTS has clearly improved faster than its benchmarks， and a system built for live agents should be judged inside live interaction， not on isolated clips.

The failure is not that speech models sound bad. It is that evaluation still treats naturalness like a single trait that can be averaged， ranked， and optimized.

That misses what listeners actually hear. A voice feels human through tiny timing shifts， restrained emotion， uneven breath， and phrasing that fits the moment rather than performs at every moment.