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

The DOI matches its Crossref record; 156 of the 156 references it cites were found in a registry or an index; 68 of the 142 sources its claims cite were checked for support.

Audio Deepfake Detection: What Has Been Achieved and What Lies Ahead

Open access
Zhang, Bowen, Cui, Hui, Nguyen, Van, Whitty, Monica · 2025 · Sensors

Abstract

Advancements in audio synthesis and manipulation technologies have reshaped applications such as personalised virtual assistants, voice cloning for creative content, and language learning tools. However, the misuse of these technologies to create audio deepfakes has raised serious concerns about security, privacy, and trust. Studies reveal that human judgement of deepfake audio is not always reliable, highlighting the urgent need for robust detection technologies to mitigate these risks. This paper provides a comprehensive survey of recent advancements in audio deepfake detection, with a focus on cutting-edge developments in the past few years. It begins by exploring the foundational methods of audio deepfake generation, including text-to-speech (TTS) and voice conversion (VC), followed by a review of datasets driving progress in the field. The survey then delves into detection approaches, covering frontend feature extraction, backend classification models, and end-to-end systems. Additionally, emerging topics such as privacy-preserving detection, explainability, and fairness are discussed. Finally, this paper identifies key challenges and outlines future directions for developing robust and scalable audio deepfake detection systems.

Overview

CoChat Summary

TL;DR

This survey reviews recent audio deepfake detection, covering generation methods (TTS and VC), datasets, frontend/backend detection approaches, evaluation metrics, and emerging topics such as privacy, fairness, explainability, anti-spoofing, and compression robustness. It concludes that detection has advanced substantially but still faces major gaps in generalisation, lightweight deployment, dataset diversity, and ethical safeguards.

Key findings

  • Human judgement of deepfake audio is unreliable: one cited study found humans correctly identified deepfake audio with only 73% accuracy, and a large-scale user study with over 1200 participants found human performance remains suboptimal in complex scenarios.
  • Detection research is shifting toward self-supervised learning (SSL) frontends such as Wav2Vec 2.0, WavLM, XLSR (53 languages), and XLS-R (128 languages), with SSL features often outperforming traditional handcrafted features like LFCC and CQCC, especially for unseen attacks.
  • Emerging defences report strong results: SafeEar forces ASR models to exceed 90% word error rate in adversarial content-recovery scenarios; AntiFake achieves over 95% protection across five speech synthesisers and three speaker verification systems; Timbre Watermarking reaches 100% extraction accuracy against preprocessing attacks and nearly 90% under extreme low-bitrate/low-pass conditions.
  • Dataset coverage is expanding but remains limited: ASVspoof editions include 106 speakers (2015), 42 speakers/177 replay sessions/61 replay configurations (2017), and 107 speakers (2019); newer datasets include MLAAD with 38 languages and 82 spoofing models, and CVoiceFake with five languages.

Methods

  • Design: comprehensive survey of recent audio deepfake detection, focusing on cutting-edge developments in the past few years.
  • Scope: reviews audio deepfake generation via text-to-speech (TTS) and voice conversion (VC), datasets, evaluation metrics (EER and t-DCF), frontend feature extraction, backend classification models, end-to-end systems, and emerging topics including privacy, fairness, adaptability, explainability, anti-spoofing, and compression robustness.
  • Analysis: thematic synthesis and comparative discussion of cited systems, with performance tables such as Table 5/Table 6 on ASVspoof systems ranked by EER and Table 7 on generalisation on the ITW dataset; no primary experimental study is conducted.

Limitations

  • The paper states that current detection models generalise poorly across languages, domains, and novel spoofing attacks, and that high-performing systems often depend on large-scale deep learning architectures unsuitable for resource-constrained settings.
  • The paper states that existing datasets lack sufficient coverage of real-world scenarios, including diverse languages, accents, background noises, and device conditions, and that imbalanced speaker-type distributions can create fairness concerns.
  • Evident from the text: as a survey, it does not provide new empirical validation, and its performance claims depend on the cited works and selected tables; the provided text also does not describe a systematic literature-search protocol.

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

Integrity signals

Registry

Crossref · 10.3390/s25071989

Cites

156 of 156 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Overview of Different audio deepfake technology: (a) text-to-speech (TTS); (b) voice conversion (VC); (c) emotion fake; (d) scene fake; (e) partially fake.
Figure 1 Overview of Different audio deepfake technology: (a) text-to-speech (TTS); (b) voice conversion (VC); (c) emotion fake; (d) scene fake; (e) partially fake.CC BY · PubMed Central
Figure 2 Overall workflow of audio deepfake detection. If the raw waveform is directly fed into the backend classification model, the model is considered an end-to-end approach.
Figure 2 Overall workflow of audio deepfake detection. If the raw waveform is directly fed into the backend classification model, the model is considered an end-to-end approach.CC BY · PubMed Central

Verification

Claims the paper makes

49 Valid · 9 Not valid · 53 Not checked · 31 Own results and interpretations

52 more claims · 23 with their cited sources checked

Create a free account to see all 142 claims and to verify the 9 references still open.

See all claims — free account
  • Probably not a substantial claim. 30
  • It may state two separate claims as one. 66

Create a free account to see these claims.

Citations

Cites

Show all 145 references

Research works printed but not linked: 15

  1. Ubert, J. (2023). Fake it: Attacking Privacy Through Exploiting Digital Assistants Using Voice Deepfakes. [Ph.D. Thesis, Marymount University].
  2. Ren, 2021, A survey on speech forgery and detection, J. Signal Process., 37, 2412
  3. Ren, Y., Hu, C., Tan, X., Qin, T., Zhao, S., Zhao, Z., and Liu, T.Y. (2020). Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv.
  4. Kong, Z., Ping, W., Huang, J., Zhao, K., and Catanzaro, B. (2020). Diffwave: A versatile diffusion model for audio synthesis. arXiv.
  5. Lee, S.G., Ping, W., Ginsburg, B., Catanzaro, B., and Yoon, S. (2022). Bigvgan: A universal neural vocoder with large-scale training. arXiv.
  6. Lee, S.H., Choi, H.Y., and Lee, S.W. (2024). Accelerating High-Fidelity Waveform Generation via Adversarial Flow Matching Optimization. arXiv.
  7. Wang, D., and Zhang, X. (2015). Thchs-30: A free chinese speech corpus. arXiv.
  8. Frank, J., and Schönherr, L. (2021). Wavefake: A data set to facilitate audio deepfake detection. arXiv.
  9. Baevski, 2020, wav2vec 2.0: A framework for self-supervised learning of speech representations, Adv. Neural Inf. Process. Syst., 33, 12449
  10. Iqbal, F., Abbasi, A., Javed, A.R., Jalil, Z., and Al-Karaki, J.N. (2022, January 17–21). Deepfake Audio Detection Via Feature Engineering And Machine Learning. Proceedings of the CIKM 2022 Workshops, Atlanta, GA, USA. CEUR Workshop Proceedings; 2022.
  11. Alexey, D. (2020). An image is worth 16 × 16 words: Transformers for image recognition at scale. arXiv.
  12. Kipf, T.N., and Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv.
  13. Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y. (2017). Graph attention networks. arXiv.
  14. Zhang, X., Yi, J., and Tao, J. (2024). EVDA: Evolving Deepfake Audio Detection Continual Learning Benchmark. arXiv.
  15. Channing, G., Sock, J., Clark, R., Torr, P., and de Witt, C.S. (2024). Toward Robust Real-World Audio Deepfake Detection: Closing the Explainability Gap. arXiv.

Cited by

See all 72

Metadata

Authors
Zhang, B., Cui, H., Nguyen, V., & Whitty, M.
Affiliations
Monash University (AU)
Corresponding author
Cui, Hui
Venue
Sensors
Publisher
MDPI AG
Published
2025-03-22
Volume
25
Issue
7
Pages
1989
ISSN
1424-8220 (ISSN-L)
OpenAlex
W4408799803
Cited by
72
References
145
Citations by year
2026: 54 · 2025: 18
FWCI
6.87
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Monash University Research Portal (Monash University) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
  • Physical Sciences › Computer Science › Signal Processing › Music and Audio Processing
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
Keywords
audio deepfake detection, deepfake audio, voice cloning, text-to-speech, voice conversion, sound synthesis, deepfake datasets, privacy-preserving detection, explainability, algorithmic fairness, sound manipulation

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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

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: Audio Deepfake Detection: What Has Been Achieved and What Lies Ahead. https://app.cochat.ai/papers/10.3390/s25071989
The paper
Zhang, B., Cui, H., Nguyen, V., & Whitty, M. (2025). Audio Deepfake Detection: What Has Been Achieved and What Lies Ahead. Sensors. https://doi.org/10.3390/s25071989

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.