Towards better knowledge distillation with multi-teacher for efficient audio deepfake detection

Abstract

This project focuses on the development and evaluation of a student model for audio deepfake detection, leveraging a single teacher model from a selection of advanced architectures, including XLSR, Facebook MMS, X-vector, UniSpeech, and Wav2Vec2. The primary objective is to enhance the detection capabilities of spoofed audio by utilizing the rich feature representations learned by these teacher models. The training process involves transferring knowledge from the teacher model to the student model, which is designed to classify audio samples as either genuine or spoofed. This approach capitalizes on the strengths of the selected teacher model, enabling the student model to learn ef fective discriminative features while maintaining a compact architecture suitable for deployment in real-world applications. To evaluate the performance of the student model, we employ the Equal Error Rate (EER) metric, which provides a balanced measure of false acceptance and false rejection rates. By analyzing EER, we can assess how well the student model generalizes to unseen spoofing attacks and its effectiveness in distinguishing between genuine and counterfeit audio. The results demonstrate that training with a single teacher model significantly enhances the student model’s performance in audio deepfake detection tasks. This study not only contributes to the ongoing research in speaker verification and spoofing countermeasures but also lays the groundwork for future explorations involving multiple teacher models and various combinations of architectures. Ultimately, this work aims to improve robustness against emerging spoofing techniques and advance the state-of-the-art in audio verification systems.