AnimeScore: A Preference-Based Dataset and Framework for Evaluating Anime-Like Speech Style
Researchers have developed AnimeScore, a new framework and dataset to objectively evaluate speech that mimics anime character voices. This system uses pairwise preference judgments, collecting 15,000 evaluations to identify key acoustic features like controlled resonance and prosodic continuity. The framework achieves up to 90.8% AUC using SSL-based ranking models, offering a practical metric for generative speech models. AI
IMPACT Provides a new objective metric for evaluating and optimizing generative speech models for niche stylistic applications.