cff-version: 1.2.0
message: "If you use VIPBench in your research, please cite the accompanying NeurIPS 2026 paper."
title: "VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning"
abstract: "VIPBench is a benchmark of 124,876 same/different identity judgments from 1,290 English-speaking listeners on 9,800 voice pairs across 100 speakers, spanning real recordings, AI voice clones, and continuously morphed voices. The release includes audio, listener judgments, and pre-extracted embeddings for ten speaker and speech-representation models. Four evaluation tasks measure model alignment with human voice-identity perception."
type: dataset
authors:
  - family-names: "Anonymous"
    given-names: "Authors"
    affiliation: "Withheld for double-blind review"
license: CC-BY-NC-4.0
version: "1.0"
date-released: "2026-05-06"
keywords:
  - speaker embeddings
  - voice identity perception
  - human-aligned benchmark
  - voice cloning
  - perceptual evaluation
preferred-citation:
  type: conference-paper
  title: "VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning"
  authors:
    - family-names: "Anonymous"
      given-names: "Authors"
  collection-title: "Advances in Neural Information Processing Systems Datasets and Benchmarks (NeurIPS Evaluations and Datasets Track)"
  year: 2026
  notes: "Anonymized for double-blind review. Author identities and DOI to be added at camera-ready."