TTS isn't solved: New research shows naturalness ≠ appropriateness across domains
Five SOTA TTS systems evaluated across 5 domains reveal major blind spots in current metrics.
A new study published at Interspeech 2026 challenges the long-held assumption that naturalness is the ultimate goal for text-to-speech (TTS) systems. Researchers from multiple institutions evaluated five state-of-the-art TTS models across five distinct domains: AI assistant, reader, actor, animated character, and spontaneous speaker. Their key finding: the perceived 'appropriateness' of a voice for a given context is largely independent of how 'natural' it sounds. Systems that scored high on conventional naturalness benchmarks still felt out of place in expressive or stylized roles, like animation or acting.
The study reveals several blind spots. While all five TTS systems performed well as readers (the standard benchmark task), their performance dropped significantly in more expressive domains. Moreover, optimizing for one domain — say, making a voice sound perfectly spontaneous — actually degraded its appropriateness for other uses, like being a calm AI assistant. Naturalness scores tended to penalize stylized, theatrical speech while rewarding spontaneous, conversational tones, suggesting current evaluation metrics are biased toward a narrow ideal. The researchers argue that TTS performance is far from 'solved' and that context-aware, multi-dimensional evaluation is essential for real-world deployment.
- Five SOTA TTS systems evaluated across 5 domains: AI assistant, reader, actor, animated character, spontaneous speaker.
- Appropriateness varies independently of naturalness – systems strong at reading, weak in expressive roles.
- Optimizing for one domain degrades others; naturalness metrics penalize stylized speech and reward spontaneity.
Why It Matters
TTS evaluation must move beyond naturalness; context-aware metrics are critical for real-world voice applications.