Reproduce “Is Natural Always Appropriate? Investigating Naturalness and Appropriateness Across Different Domains for TTS Evaluation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dominika Woszczyk, Andreas Triantafyllopoulos, Jura Miniota, Éva Székely, Bjoern Schuller
An investigation into Text-to-Speech (TTS) evaluation reveals that listener judgments of speech appropriateness are strongly context-dependent across diverse domains, often decoupling from perceived human-likeness, and identifies that common automatic metrics are frequently inadequate for expressive contexts.
Key insights
- Perceived appropriateness is highly domain-dependent, with specific acoustic features (e.g., rhythmic variability for animated characters, neutral stability for AI assistants) preferred for different contexts.
- A 'human-likeness paradox' exists, where a highly human-like voice may be inappropriate for certain tasks (e.g., AI assistant) and vice-versa, indicating a complex, often decoupled, relationship between these two evaluation dimensions.
- Many common automatic quality estimators (e.g., UTMOS, DNSMOS) show negative correlations with appropriateness in expressive or spontaneous domains, highlighting their unsuitability for diverse TTS evaluation contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.