Skip to content

Dataset · datapointai

315 thousand human evaluations for 15 text-to-speech models

datapointai/text-to-speech-human-preferences-315k

Pairwise audio comparisons from 15 TTS models in English for RLHF and quality evaluation.

What's inside

The dataset contains 315,000 pairwise user votes comparing 15 text-to-speech models across 300 English prompts from eight voice agent categories. The set includes all generated audio recordings and tabular evaluation data. The dataset is gated and underlies the Datapoint Audio Bench benchmark.

What you can do with it

  • Aligning text-to-speech models with human preferences (RLHF)
  • Training reward models for automated voice quality evaluation
  • Comparing and benchmarking voice assistant text-to-speech systems
Tasks
Text → speech, Reinforcement learning
Languages
English
Size
100K–1M rows
Downloads this month
226
Likes
42
Access
on request on Hugging Face

How to load it

from datasets import load_dataset

ds = load_dataset("datapointai/text-to-speech-human-preferences-315k")