Year: 2025
Type: conference-paper
Abstract: With the rapid advancement of computer technology, especially high performance computing and deep learning algorithms, speaker verification systems have become increasingly critical for secure voice-based authentication. To counter replay and synthetic spoofing attacks in speaker verification systems, this paper introduces a lightweight residual convolutional network for robust anti‑spoofing. We extract log‑power mel‑spectrograms and phase‑based group delay features to capture both amplitude and fine‑structure characteristics of the speech signal. These features are processed through a compact residual backbone augmented with self‑attention to model long‑range temporal dependencies. A statistical fusion mechanism aggregates frame‑level representations into an utterance‑level score, which is classified by a calibrated lightweight classifier. Experiments on the ASVspoof 2019 dataset demonstrate an equal‑error rate (EER) of 0.89%, representing a significant relative improvement over the baseline. The proposed approach operates in real time on commodity hardware and generalizes well to unseen attack types, making it suitable for next‑generation voice authentication applications.
Authors: Kunming Wang, Xianchang Fan, Ying Lei
Institution: Guangdong Polytechnic of Industry and Commerce
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4414701485