COIN-AT-PVAD: A Conditional Intermediate Attention PVAD

Personalized voice activity detection (PVAD), compared to conventional VAD, shows more developmental potential in scenarios with multiple speaker interference. Among the various methods for integrating speaker and acoustic features, performance may be limited due to the weaker representational capability of speaker embeddings derived from external speaker verification models. This study proposes a new architecture called Conditional Intermediate Attention PVAD (COIN-AT-PVAD) to address this issue. This architecture builds upon the Attentive Score (AS) module and incorporates the Feature-wise Linear Modulation (FiLM) scheme to better integrate multimodal information. Through comparing various fusion strategies, we show that COIN-AT-PVAD significantly surpasses the baseline model, especially when external embedding features have limited representational capacity. Experimental findings also indicate that, when compared to some state-of-the-art models, COIN-AT-PVAD achieves superior average precision and accuracy while retaining a compact model size, showcasing its efficacy in real-world applications on resource-limited devices.