mueller91/human-perception-audio-deepfake-2026

Human Audio Deepfake Perception 2026

A large-scale listening study evaluating how well humans detect modern audio deepfakes. The dataset contains 35,532 deepfake-detection judgments from 1,768 anonymous participants across 138 TTS and voice-conversion systems, collected via a publicly accessible online listening game in 2025–2026.

This is the successor to the 2021 ASVspoof-2019 perception study (Müller, Pizzi & Williams, 2022) and extends the same paradigm to modern systems, including commercial APIs (ElevenLabs, Resemble AI, Cartesia), autoregressive LM-based TTS (VALL-E, Bark, ChatTTS, Llasa, etc.), flow-matching systems (F5-TTS, CosyVoice), and others.

Headline findings

• Skepticism shift. Human accuracy on fake samples is essentially unchanged from 2021 (72.9% → 71.2%), but accuracy on real audio dropped sharply (72.7% → 64.1%). Listeners increasingly misclassify authentic speech as fake.
• Hardest architectures. Commercial APIs (61.3%) and AR-LM systems (66.0%) produce the hardest-to-detect samples; classical seq2seq (75.4%) and flow-matching models (76.8%) remain easier.
• ML reference. A Wav2Vec 2.0 + AASIST detector maintains 94.5% overall accuracy across all categories.

See the accompanying paper for full results and discussion.

Dataset structure

One row = one round = one (participant, audio sample) judgment.

field	type	description
uid	str	SHA-256-hashed (salted) anonymous participant id, 16 hex chars
rounds_played	int	1-indexed round number for this participant
filename	str	bare audio filename (no path)
attack_id	str	TTS/VC system, or one of - / InTheWild / LJSpeech / ASVSpoof5_- for real audio
true_label	str	real or fake
user_decision	str	participant's classification: real or fake
ml_decision	str	reference ML detector's prediction: real or fake

Demographic attributes (age bracket, IT skill 1–5, native English) were collected during the study but are excluded from release to prevent re-identification from response patterns.

Loading

from datasets import load_dataset

ds = load_dataset("mueller91/human-audio-deepfake-perception-2026")
print(ds["train"][0])

Or directly with pandas:

import pandas as pd
df = pd.read_csv(
    "hf://datasets/mueller91/human-audio-deepfake-perception-2026/data.csv"
)

Getting the audio

The CSV references filenames only. Audio lives in four public corpora; join on filename to attach audio to each judgment:

source	content	location
MLAAD (English subset)	majority of fake samples	mueller91/MLAAD
ASVspoof 5	additional fake + some real	https://www.asvspoof.org/
In-The-Wild	real samples	https://deepfake-detection.com/in-the-wild-dataset
LJSpeech	real samples	https://keithito.com/LJ-Speech-Dataset/

Architecture groupings

The 138 systems are grouped into 10 architecture families used in the paper:

• Seq2Seq: encoder-decoder with attention (e.g. Tacotron 2)
• VITS: VAE + flow + GAN
• XTTS: GPT-based multi-speaker TTS with VITS decoder
• Flow: flow-matching models (F5-TTS, CosyVoice)
• Diffusion: diffusion-based (Grad-TTS, StyleTTS 2)
• AR-LM: autoregressive LM over codec tokens (VALL-E, Bark, ChatTTS, Llasa, MOSS-TTS, Sesame CSM, Kani-TTS, LFM2.5-Audio, …)
• VC: voice conversion (RVC, OpenVoice V2)
• Commercial: proprietary APIs (ElevenLabs, Resemble AI, Cartesia, …)
• ASVSpoof5: attacks from the ASVspoof 5 challenge
• Other: uncategorized

Anonymization

• Participant IDs are SHA-256 hashes of internal session UIDs with a project-specific salt; 16 hex characters retained. Original UIDs are not derivable.
• Demographic attributes are excluded.
• Self-selected web volunteers; no directly identifying information (names, emails, IP addresses) was ever collected.

Limitations

• English-only.
• Self-selected sample skews younger.
• Audio quality varies with participants' playback equipment and browser compression.
• Active-learning sampling produces uneven per-attack sample counts; attacks with fewer than 10 judgments are excluded.
• Participation is open and anonymous, so we cannot control for users who may have participated in both the 2021 and the 2026 studies.

Citation

@misc{mueller2026erodingtrust,
  title  = {Eroding Trust in Real Speech:
            A Large-Scale Study of Human Audio Deepfake Perception},
  author = {M\"uller, Nicolas M. and Choong, Wei Herng},
  year   = {2026},
  note   = {Preprint forthcoming}
}

Predecessor study (cite for the 2021 baseline):

@inproceedings{muller2022human,
  title     = {Human Perception of Audio Deepfakes},
  author    = {M\"uller, Nicolas M. and Pizzi, Karla and Williams, Jennifer},
  booktitle = {Proc. 1st International Workshop on Deepfake Detection
               for Audio Multimedia (DDAM)},
  pages     = {85--91},
  year      = {2022},
  doi       = {10.1145/3552466.3556531}
}

License

CC-BY-NC-4.0. Free for research and non-commercial use with attribution.