Human Audio Deepfake Perception 2026: 35,532 Detection Judgments

A large-scale listening study collected 35,532 deepfake-detection judgments from 1,768 anonymous participants. The data was gathered via an online listening game in 2025–2026, evaluating 138 TTS and voice-conversion systems. It is the successor to the 2021 ASVspoof-2019 perception study, authored by mueller91.

Use Cases

• Benchmarking human detection performance against automated systems based on the described listening study paradigm.
• Analyzing trends in human perception of audio deepfakes across 138 different TTS and voice-conversion systems.
• Training or calibrating automated deepfake detection models using human judgment data as a reference.
• Studying the evolution of deepfake quality and human discernment by comparing with the earlier 2021 study.

Strengths

• Large-scale collection with 35,532 individual human judgments.
• Broad system coverage, evaluating 138 different TTS and voice-conversion systems.
• Involves a substantial participant pool of 1,768 anonymous individuals.
• Builds upon a known prior study (ASVspoof-2019 perception study), suggesting a consistent methodology.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect temporal and platform bias inherent to the specific online listening game and participant pool.

Provenance

Source

huggingface user mueller91

Collection Method

Collected via a publicly accessible online listening game.

Time Range

Data collection occurred in 2025–2026.

Freshness

Last updated 2026-05-21 09:39:38