An Approach Combining CNN and VGG16 for Detecting Voice Spoofing

Résumé fourni par la source

Voice Authentication is essential for protecting private devices and financial activities. The challenges to precision and resilience in voice authentication are to detect and achieve due to the rapid progress of artificial intelligences techniques which used to deep fake voice and variation spoofing speech Recent advances in the deep learning industry, especially transfer learning of VGG16 based on convolutional neural networks (CNN), provide an opportunity to improve voice recognition mechanisms. This paper uses advanced deep transfer learning, such as VGG16, and an adapted CNN to enhance the security and accuracy of authentication voice. The aim is to evaluate and analyze the accuracy of different approaches for authenticating voice recognition. Whereas the dataset is 1.5 GB and is utilized to assess the VGG16 and a developed CNN. The assessment includes 17,870 voice samples, with an even allocation of 1088 samples from fake and real contributors in testing, 13,965 samples for fake and real contributors in training, and 2,826 for fake and real contributors in validation. The models were assessed based on their ability to recognize and authenticate samples' voices. The VGG16 and CNN models show the best level of accuracy in recognizing the voice, with an accuracy rate approximately 99.83%. CNN provided sufficient performance at approximately 98.89%. Furthermore, the VGG16 demonstrated a higher accuracy. The model is efficient for voice authentication technique evaluated. The article maintains that using transformers outperforms CNNs when it comes to voice recognition. They determine context more accurately and handle noise more appropriately.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.