Skip to content

Dataset · sendfuze

Human voice and deepfake perception benchmark

sendfuze/vipbench

About 125K human voice similarity ratings on pairs of real recordings, AI clones, and morphs for speaker recognition testing.

What's inside

The collection includes 124,876 ratings ("same or different person") from 1,290 English-speaking listeners for 9,800 audio pairs from 100 speakers. Pairs include three recording types: real voices, AI clones from a modern TTS model, and smoothly morphed voices. Besides audio, the dataset contains tabular data and English text.

What you can do with it

  • Evaluating speaker-embedding model robustness against modern TTS voice clones
  • Calibrating audio deepfake detectors to human perception
Tasks
Audio classification, Other
Languages
English
Size
100K–1M rows
Downloads this month
613
Likes
2

How to load it

from datasets import load_dataset

ds = load_dataset("sendfuze/vipbench")