ai
Researchers Propose Audio Spoofing Detection System for English and Arabic
A paper published on May 9, 2026, introduces Phonetic-DeepKANet, a dual-modality framework designed to detect audio deepfakes and other spoofing attacks in English and Arabic. The system combines deep feature extraction with acoustic-phonetic representations and a Kolmogorov Arnold Network classifier. The work also presents a new Arabic audio spoofing dataset to address gaps in language coverage.
May 8, 8:00 PM·2m read1 source
thehindu.comAudio version
Tap play to generate a narrated version.
Developing·Limited corroboration so far. This page will refresh as more sources emerge.
A research paper published today describes a new framework for detecting audio spoofing attacks, including deepfakes, that target automatic speaker verification systems. The framework, named Phonetic-DeepKANet, was developed to improve generalization against unknown attacks and to extend coverage to Arabic, a language often overlooked in such research due to limited datasets.
The approach uses a dual-modality design that pairs deep features extracted by TransRawNet with multi-view acoustic-phonetic representations. These combined features are then processed by a Kolmogorov Arnold Network classifier. The paper states that this combination allows reliable detection across voice conversion, text-to-speech, partial spoofing and codec-related distortions.
Researchers introduced an Arabic audio spoofing dataset as part of the work to support evaluation in that language. The system was tested on the ASVspoof-2019 LA, ASVspoof-2021 LA, ASVspoof-2021 DF, partial spoofing and the newly created Arabic dataset.
Experiments included algorithm-wise and cross-corpora evaluations.
The framework recorded a minimum tandem detection cost function of 0.09 on the ASVspoof-2019 LA dataset and 0.14 on the ASVspoof-2021 LA dataset. On the Arabic dataset it achieved an equal error rate of 8.06 percent. According to the paper, the method produced the best result for logical access attacks among 41 entries in the ASVspoof-2021 challenge and the third-best equal error rate of 17.55 percent on the deepfake track among 33 participants.
