Authors:
Bowen Zhang, Hui Cui, Van Nguyen, and Monica Whitty
Where published:
Sensors, 2025, Volume 25, Article 1989.
Abstract:
This study is essentially a comprehensive survey paper on audio deepfake detection. That means it doesn’t introduce a single new model – instead, it organizes, analyzes, and critiques the entire research landscape so others (like you) can understand where the field stands and where to go next.
Dataset names (used for):
- The paper reviews datasets used for training, validation, benchmarking, and evaluating generalization in audio deepfake detection. Main datasets include ASVspoof, ADD Challenge, FoR, WaveFake, HAD, ITW, LibriSeVoc, SceneFake, EmoFake, CVoiceFake, and MLAAD.
Some description of the approach:
The paper provides a comprehensive survey of audio deepfake detection. It reviews audio deepfake generation methods such as TTS and VC, discusses datasets, evaluation metrics, frontend feature extraction, backend classification models, end-to-end systems, and emerging issues such as privacy, fairness, adaptability, explainability, and robustness.
Keywords:
audio deepfake detection; text to speech (TTS); voice conversion (VC).
Subject Area:
Audio deepfake detection, speech synthesis, voice conversion, automatic speaker verification, cybersecurity, biometric authentication, digital media trust, and sensor/IoT security.
Associated Tools:
- Reviewed generation technologies include TTS and VC models such as DeepVoice, Tacotron, FastSpeech, CycleGAN-VC, AutoVC, MulliVC, VQVC, and FreeVC.
- Detection-related tools/models/features include GMM, SVM, LCNN, ResNet, SENet, CapsNet, ViT, GCN/GAT, RawNet/RawNet2, AASIST, W2V2, XLS-R, WavLM, SLIM, SafeEar, and others.
Feature Type:
The paper groups features into handcrafted features and learning-based features. Handcrafted examples include LPS, LFB, Mel Spectrogram, CQT, Spec, MFCC, LFCC, and CQCC. Learning-based examples include SincNet, RawNet/RawNet2, Wav2Vec 2.0, XLSR/XLS-R, WavLM, SLIM, and VGG-based spectrogram features.