Beyond Naturalness: Probing Automated Text-To-Speech Evaluators on Linguistically Grounded Dimensions

Automated Text-to-Speech (TTS) evaluation methods (Mean Opinion Score (MOS) predictors and Audio Large Language Models (Audio-LLM) judges) are expected to reflect human perception, yet it is unclear how well they capture the distinct aspects of speech that listeners actually perceive. We deconstruct "naturalness" into a linguistically grounded annotation schema spanning 10 distinct perceptual dimensions, and use it to construct the first dimension-level meta-evaluation benchmark for TTS, comprising 860 utterances annotated by trained linguist raters. Results from benchmarking four MOS predictors and four Audio-LLM judges reveal that MOS predictors collapse onto acoustic signal quality, while Audio-LLM judges show selective, prompt-dependent detection that does not generalise across all dimensions. Neither class reliably captures a breadth of linguistically structured speech errors. Our dataset, annotation schema, and evaluation code are publicly released to support more targeted and interpretable TTS evaluation.

Is Natural Always Appropriate? Investigating Naturalness and Appropriateness Across Different Domains for TTS Evaluation

Text-to-speech (TTS) evaluation is an open challenge. While the primary target was "naturalness," recent fidelity gains shifted focus toward "appropriateness" and whether speech is correct for its context. In this work, we examine how perception changes when the expected downstream use varies. We measure the appropriateness and human-likeness of five SOTA TTS systems across five domains: AI assistant, reader, actor, animated character, and spontaneous speaker. Results show appropriateness varies across domains independently of naturalness. While systems shine at reading, expressive domains remain challenging, and optimizing for one can degrade others. Furthermore, naturalness scores tend to penalize stylized speech while rewarding spontaneity. Finally, our study also highlights blind spots in one-size-fits-all evaluation metrics across more expressive domains. We demonstrate that TTS performance is not "solved" but depends on the target domain, requiring context-aware evaluation.

Naturalness Predicts but Does Not Cause Transferability in Image Encodings of Real-World Streams

A common practice converts a one-dimensional signal into an image so that a vision backbone pretrained on natural photographs can be reused for recognition, yet the encoded image is rarely examined. We ask how the visual naturalness of an encoded image relates to its transfer accuracy under a frozen backbone. We build WorldStream, a corpus of 299 heterogeneous current-value series from key-free public APIs (weather, air quality, earthquakes, gold and oil, equities, crypto, foreign exchange, web activity and space weather), with a nine-way source-recognition task over 3143 temporally split windows. Across seven encodings and six frozen backbones, the Frechet distance of an encoding to natural images (FID) predicts its accuracy: Spearman $ρ=-0.72$. Two controlled interventions show this is not causal in the spectrum. Our invertible encoder has a single adjustable part, a spectral exponent $β$ (power $\propto |f|^{-β}$); varying $β$ moves the image toward or away from the natural-image manifold at fixed content. FID is lowest near the natural value $β\approx 2$, but frozen accuracy stays flat and far below the structured baselines (19.2% vs. 73.0%), and FID and accuracy are only weakly related over the sweep (Pearson $-0.32$). A second intervention, phase scrambling, holds the power spectrum exactly fixed while removing local structure; now FID and accuracy fall together (Pearson $-0.89$). The cross-encoding correlation is thus mediated by local structure, not spectral naturalness: FID predicts accuracy because Inception reads the same structure the backbones do. Full fine-tuning does not close the gap (27% vs. 67%), so the deficit is structural. The encoder is exactly invertible, recovering the signal from the 8-bit image at 72.9 dB, so the image doubles as a lossless record of the data.