Гібридна система детекції синтезованого мовлення на основі SSL-ембеддінгів

Abstract

This paper presents a hybrid system for synthetic speech detection that integrates spectral, phase-based, and prosodic acoustic features with self-supervised learning (SSL) embeddings extracted from the WavLM model. The proposed multimodal approach combines low-level handcrafted features (LFCC, jitter, shimmer, phase stability) with high-level context-aware representations derived from intermediate transformer layers. Such integration enables the model to capture both fine-grained acoustic artifacts and global structural inconsistencies in synthetic speech. The proposed architecture shows strong potential for applications in biometric security and digital media integrity.