Beyond Known Speakers: Enhancing Spoofing Detection Generalization with VIVED

Abstract

Recent speech synthesis systems have moved beyond conventional voice cloning toward the generation of expressive, text-described virtual identities. This shift creates a generalization challenge for spoofing countermeasures trained primarily on cloned speech from a limited pool of reference speakers and relatively neutral read-speech content. We introduce Virtual Identities with Varied Emotional Dynamics (VIVED), a large-scale dataset designed to improve robustness against text-driven virtual voices. VIVED is built around three principles. First, it generates over 100,000 synthetic virtual identities from stochastic natural-language descriptions sampled from a seven-dimensional acoustic and paralinguistic attribute bank, reducing reliance on real speaker reference audio for spoof generation. Second, it combines linguistically and affectively diverse transcripts from LibriSpeech, TIMIT, and the EmoBox English subset to expose synthesis models to phonetic complexity and emotional prosody. Third, it applies a symmetric domain-matching and acoustic laundering pipeline to both bona fide and synthetic utterances, including multi-codec alignment, room impulse response simulation, noise injection, and codec compression, thereby reducing shortcut learning from channel or processing artifacts. Empirical analysis indicates that VIVED covers a speaker representation space with nearly 1.8 times the Vendi Score of the ASVspoof 5 baseline, substantially reducing the risk of identity memorization. Extensive evaluations demonstrate that training on VIVED systematically enhances robustness to text-driven virtual voices while maintaining competitive cross-corpus performance on ASVspoof 5. Multiple baseline architectures achieve substantially lower EER and minDCF on the VIVED evaluation set compared with models trained only on ASVspoof 5.