A 200-join stress test for expressive TTS

Acting Transitions packages 200 part-to-part transitions from synthetic performances, each rated from 0 to 5 for how naturally two generated pieces join. The Apache-2.0 dataset includes part A, part B, their joined audio, VoiceCLAP embeddings, and discontinuity features for every sample, while explicitly warning that its labels come from one model judge rather than human listeners. Voice builders extending audio-first interfaces beyond transcription can use it to test a neglected failure mode—whether an intended emotional turn sounds continuous rather than edited—but should calibrate the automated scores against human listening before adopting them as a release gate.