ServiceNow-AI/Beyond-Naturalness-TTS-Eval

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

Dataset Overview

• Purpose: Evaluate automated TTS evaluators, including MOS predictors and Audio-LLM judges.
• Data: 860 synthetic speech samples covering ten linguistically grounded dimensions.
• Labels: Majority votes from three professional linguists; 0 = failure, 1 = no failure.
• Coverage: Every sample retains all ten labels, including dimensions other than its generation target.
• Use: Evaluation only; no model training or fine-tuning. See usage restrictions.

Loading the Dataset

from datasets import load_dataset

repo = "ServiceNow-AI/Beyond-Naturalness-TTS-Eval"

# Full benchmark: 860 samples. This is also the default configuration.
all_samples = load_dataset(repo, "all", split="test")

# Targeted evaluation: with all ten human-label columns retained.
phonetic = load_dataset(repo, "phonetic_accuracy", split="test")
gold_labels = phonetic["Phonetic Accuracy"]  # 0 = failure, 1 = no failure

• all is the default configuration; dimension subsets group rows by Target Dimension.
• Every configuration has one test split and retains all ten label columns.

Subsets and Dimensions

• Each dimension subset is balanced 50/50 on its target label. Other labels need not be balanced.
• Counts and Krippendorff's α per dimension are mentioned below:

Subset	Human-label column	Failure condition	Samples	α
phonetic_accuracy	Phonetic Accuracy	Sounds outside the acceptable realization of a lexical target	110	0.767
lexical_stress	Lexical Stress	Primary stress on the wrong syllable	90	0.821
intonation	Intonation	Pitch contour inappropriate to utterance type or discourse function	80	0.469
prosodic_stress	Prosodic Stress	Word or phrase prominence inconsistent with information structure	102	0.658
prosodic_boundary_placement	Prosodic Boundary Placement	Prosodic boundaries inconsistent with syntactic structure	88	0.727
speech_rate_appropriateness	Speech Rate Appropriateness	Delivery pace inappropriate to context	128	0.707
emotional_appropriateness	Emotional Appropriateness	Emotional tone inconsistent with utterance content	46	0.596
expressiveness	Expressiveness	Insufficient or excessive variation in delivery and vocal energy	94	0.460
speaker_identity_consistency	Speaker Identity Consistency	Perceived speaker characteristics change within an utterance	66	0.713
human_plausibility	Human Plausibility	Vocal qualities outside the range of human production	56	0.710
all	All ten columns	Combined benchmark	860	—

• Word: Phonetic Accuracy, Lexical Stress.
• Prosodic: Intonation, Prosodic Stress, Prosodic Boundary Placement, Speech Rate Appropriateness.
• Paralinguistic: Emotional Appropriateness, Expressiveness, Speaker Identity Consistency, Human Plausibility.

Schema

All configurations share these columns:

Column	Type	Description
Unique ID	string	Sample identifier, e.g. 1-1002.
Target Dimension	string	Generation target; determines subset membership.
Source Dataset	string	Source/provenance label.
Original Text	string	Reference text before manipulation.
Audio	audio	Audio decoded at 44,100 Hz.
Error Description	string	Intended manipulation, not a human label.
Synthesized Sentence	string	Text or phoneme string sent to the TTS engine.
Target Dimension Rater Scores	list[int64]	Three individual binary ratings for Target Dimension, e.g. [0, 1, 1], before majority voting; earlier ternary ratings are collapsed.
Ten human-label columns listed above	int64	Majority-voted labels: 0 = failure, 1 = no failure.

Dataset Construction

• Sources: Harvard Sentences, EmergentTTS-Eval, and our own synthetically generated text.
• Error generation: Modified IPA, modified text, mismatched emotion tags, and acoustic manipulation.
• Annotation: Three linguists, with guidelines refined through a pilot experiment.
• Finalization: Majority voting, then 50/50 balancing within each targeted subset, prioritizing consensus annotations.

Limitations

• Controlled errors: Results on injected failures may not generalize to natural deployment errors.
• Limited coverage: English data, often short sentences and thus other languages, dialects, and long-form speech are not established use cases.
• Synthesis diversity: IPA-controlled samples use one TTS architecture, Cartesia Sonic 3.5.
• Label granularity: Binary labels indicate failure, not severity.
• Uncertainty: Subsets contain 46–128 samples; annotator agreement varies. Report uncertainty.
• Cross-dimensional effects: One manipulation can affect several dimensions. Balanced subsets do not reflect real-world failure rates.

Citation

Please cite the paper when using this dataset:

@article{bamgbose2026beyond,
  title   = {Beyond Naturalness: Probing Automated Text-To-Speech Evaluators on Linguistically Grounded Dimensions},
  author  = {Bamgbose, Oluwanifemi and Shah, Jash and Rosen, Simon and Brin, Lindsay Devon and Nguyen, Hoang H. and Koelzer, Anke and Hansen, Rachel and Bogavelli, Tara and Riols, Fanny},
  journal = {arXiv preprint arXiv:2608.09930},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.09930}
}

License and Usage Restrictions

• Evaluation only. Training or fine-tuning models on this dataset is not permitted. Generated audio is subject to the TTS providers' usage restrictions.
• ElevenLabs: Its Prohibited Use Policy restricts using outputs for model training, fine-tuning, or improvement, including training datasets, unless expressly permitted under the applicable terms.
• Cartesia: Its Terms of Service prohibit using outputs to develop competing products, models, or services.
• Source materials: Harvard Sentences originate from IEEE Std 269; EmergentTTS-Eval uses Apache 2.0. See the paper's artifact-license appendix.