Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 41 of the 41 references it cites were found in a registry or an index; none of the 11 sources its claims cite is checked for support yet.

Does Audio Deepfake Detection Generalize?

Open access
Müller, Nicolas, Czempin, Pavel, Diekmann, Franziska, Froghyar, Adam, Böttinger, Konstantin · 2022 · Interspeech 2022

Abstract

Current text-to-speech algorithms produce realistic fakes of human voices, making deepfake detection a much-needed area of research. While researchers have presented various techniques for detecting audio spoofs, it is often unclear exactly why these architectures are successful: Preprocessing steps, hyperparameter settings, and the degree of fine-tuning are not consistent across related work. Which factors contribute to success, and which are accidental? In this work, we address this problem: We systematize audio spoofing detection by re-implementing and uniformly evaluating architectures from related work. We identify overarching features for successful audio deepfake detection, such as using cqtspec or logspec features instead of melspec features, which improves performance by 37% EER on average, all other factors constant. Additionally, we evaluate generalization capabilities: We collect and publish a new dataset consisting of 37.9 hours of found audio recordings of celebrities and politicians, of which 17.2 hours are deepfakes. We find that related work performs poorly on such real-world data (performance degradation of up to one thousand percent). This may suggest that the community has tailored its solutions too closely to the prevailing ASVSpoof benchmark and that deepfakes are much harder to detect outside the lab than previously thought.

Overview

CoChat Summary

TL;DR

This paper re-implements and uniformly evaluates twelve audio deepfake detection architectures, finding that feature choice and input length strongly affect performance. It also introduces a new 37.9-hour “in-the-wild” dataset and shows that models trained on ASVspoof 2019 generalize poorly to real-world deepfakes, with EER degradation up to 1000%.

Key findings

  • Raw-waveform models outperform spectrogram-based models: up to 1.2% EER on ASVspoof and 33.9% EER on the “in-the-wild” dataset (RawGAT-ST and RawNet2), versus up to 6.3% EER on ASVspoof and 37.4% on “in-the-wild” for spectrogram-based models (LCNN and MesoNet).
  • Replacing melspec with cqtspec or logspec improves performance by 37% EER on average, all other factors constant.
  • Full-length audio input outperforms fixed 4-second input: average EER on ASVspoof drops from 19.89% to 9.85% when full-length input is used.
  • On the new “in-the-wild” dataset, EER values deteriorate by about 200 to 1000 percent compared with ASVspoof 2019, and models often perform no better than random guessing; training RawNet2 with 4s input on all ASVspoof 2019 splits yields 33.1 ± 0.2% EER, no improvement over training with only train and dev data.

Methods

  • The authors re-implemented twelve architectures from related work and evaluated them under a common standard, training on the ASVspoof 2019 Logical Access train and dev splits and testing on the ASVspoof 2019 eval split and a new “in-the-wild” dataset.
  • The new dataset contains 37.9 hours of audio: 17.2 hours of deepfakes and 20.7 hours of authentic material from 58 English-speaking celebrities and politicians; fake clips were created by segmenting 219 publicly available video and audio files that explicitly advertise audio deepfakes, and authentic clips were manually collected from public sources. Clips average 4.3 seconds and were downsampled to 16 kHz.
  • Models using preprocessed features were evaluated on 513-dimensional cqtspec, logspec, and melspec features; other models used raw waveforms. Experiments compared fixed 4-second inputs with variable-length inputs, were run three times with random initialization, and report averaged EER and t-DCF with standard deviation; t-DCF is reported only for ASVspoof.

Limitations

  • Stated in the paper: the authors do not fine-tune hyperparameters, which likely explains why their ASVspoof implementations perform about 2–4% EER and 0.1 t-DCF worse than related work. t-DCF is not reported for the “in-the-wild” dataset because false alarm and miss costs are available only for ASVspoof. The new dataset is intended as evaluation data.
  • Stated in the paper: all ASVspoof splits are fundamentally based on the same VCTK dataset, so adding the eval split does not improve real-world generalization.
  • Evident from the text: the “in-the-wild” dataset is limited to English-speaking celebrities and politicians, uses fake clips from publicly advertised deepfakes, relies on manual collection of authentic counterparts, and consists of short clips averaging 4.3 seconds downsampled to 16 kHz; evaluation covers ASVspoof 2019 Logical Access and this new dataset only.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.21437/interspeech.2022-108

Cites

41 of 41 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 0 Not valid · 8 Not checked · 35 Own results and interpretations

22 more claims

Create a free account to see all 43 claims and every source verdict.

See all claims — free account
  • The quote only partly states this claim. 2
  • It may state two separate claims as one. 9
  • Probably not a substantial claim. 8

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 29 references

Research works printed but not linked: 1

  1. A. Nautsch, X. Wang, N. Evans, T. H. Kinnunen, V. Vestman, M. Todisco, H. Delgado, M. Sahidullah, J. Yamagishi, and K. A. Lee, "ASVspoof 2019: Spoofing Countermeasures for the De- tection of Synthesized, Converted and Replayed Speech," vol. 3, no. 2, pp. 252-265.

Cited by

See all 258

Metadata

Authors
Müller, N., Czempin, P., Diekmann, F., Froghyar, A., & Böttinger, K.
Affiliations
  • Fraunhofer Institute for Applied and Integrated Security (DE)
  • Technical University of Munich (DE)
Corresponding author
Müller, Nicolas
Venue
Interspeech 2022
Publisher
ISCA
Published
2022-09-18
Pages
2783-2787
ISSN
2308-457X (ISSN-L)
OpenAlex
W4297841787
Cited by
258
References
29
Citations by year
2026: 82 · 2025: 93 · 2024: 60 · 2023: 19 · 2022: 4
FWCI
24.31
Type
conference paper
Language
en
Open access
Locations
  • Interspeech (USB) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • Fraunhofer-Publica (Fraunhofer-Gesellschaft) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Signal Processing › Music and Audio Processing
Keywords
audio deepfake detection, generalization, synthetic speech, audio forensics

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Does Audio Deepfake Detection Generalize? https://app.cochat.ai/papers/10.21437/interspeech.2022-108
The paper
Müller, N., Czempin, P., Diekmann, F., Froghyar, A., & Böttinger, K. (2022). Does Audio Deepfake Detection Generalize? Interspeech 2022. https://doi.org/10.21437/interspeech.2022-108

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.