Human Preference Data For TTS Audio Quality Evaluation

TTS Human Preferences (Medium) is a dataset for text-to-speech audio quality evaluation. It contains 2,000 rows, each with two TTS audio renderings and 15 human preference annotations, totaling 30,000 annotations. The dataset was created by datapointai and last updated in March 2026.

Use Cases

• Fine-tuning TTS models via direct preference optimization using the 15 human preference annotations per audio pair.
• Training a reward model for RLHF by learning to predict human preferences between two TTS audio renderings.
• Benchmarking the naturalness of different TTS systems using the aggregated human preference scores.
• Analyzing the correlation between specific audio features and human judgments of naturalness within the annotated pairs.

Strengths

• Contains 30,000 human preference annotations across 2,000 data points.
• Each data point provides a direct A/B comparison of two TTS audio outputs for the same text prompt.

Limitations

• The dataset is a medium subset of 2,000 rows, which may be insufficient for training large models without augmentation.
• Potential bias exists as annotations are sourced from a single region (US), limiting geographic diversity in preference judgments.

Provenance

Source

datapointai via Hugging Face.

Collection Method

Human annotations collected for pairs of TTS audio renderings.

Freshness

Last updated March 2026.

Geography

United States (based on platform tag 'Regionus').