Local Self-Attention based Hybrid Multiple Instance Learning for Partial Spoof Speech Detection

The development of speech synthesis technology has increased the attention towards threat spoofed speech. Although various high-performance spoofing countermeasures (CMs) have been proposed in recent years, a particular scenario is overlooked: partially-spoofed audio, where utterances may contain both and bona fide segments. Currently, research on detection lacking. existing methods either train with at utterance level, resulting gradient conflicting segment or directly level data, which requires labels that are difficult to obtain practice. In this study, better detect when only available, we formulate into multiple instance learning (MIL) problem. typical MIL uses pooling layer fuse patch scores as whole, propose hybrid (H-MIL) framework based max log-sum-exp (LSE) methods, can learn representations improve performance. Theoretical experimental verification shows H-MIL effectively relieve vanishing problems. Besides, analyze local correlations between segments introduce self-attention mechanism enhance features, further promotes our experiments, provide not results levels, but also some detailed visualization analysis, including effect spoof ratio cross-dataset detection. demonstrate effective performance method especially dealing low attacks. confirm approach deal than previous methods.