speech generation

SpeechParaling-Bench tests speech models on nuance

A new benchmark expands paralinguistic speech evaluation past coarse labels, using 1,000+ queries and pairwise judging to expose model gaps.