Меню
Vol. 2 No. 31 (2026) / 2026-06-30 / Number of views: 99
Authors
Keywords
Link to DOI:
How to quote
The article represents a systematical analytical review of the current methods in the detection of synthesized speech and protection against voice spoofing attacks of automatic speaker verification systems. The key types of voice spoofing attacks are discussed, such as text-to-speech synthesis, voice transformation, playback attacks, audio deepfakes, and their hybrid situations.
There is a comparison of the datasets used, acoustic properties, model design, training regimes, and performance measures. It is known that the current state of research is linked with the commonness of the ASVspoof datasets and supervised learning methods, which assure the comparability of the findings, but limit the applicability of models in the situation of cross-dataset transfer and domain shift. They show that the key unsolved problems are: resistance to previously unknown types of attacks, noise sensitivity and channel distortion sensitivity and poor interpretability of existing models.
Based on the analysis, the main methodological limitations were identified, and priority areas for further research were formulated, related to the development of cross-domain, reliable, and explainable detection systems intended for practical integration into ASV protection schemes. The paper also proposes a classification of synthetic speech detection methods that takes into account the type of features used and the architectural characteristics of the models.