HySV: A Hymba-Inspired Hybrid-Head Framework for Quality-Aware and Deployment-Aware Speaker Verification in Intelligent Embedded Systems

초록

Speaker verification is an important biometric technology for secure and personalized human-computer interaction in intelligent embedded systems. However, deploying deep speaker verification models on edge devices remains challenging because of restricted computational resources and strict real-time latency requirements. Existing systems commonly rely on convolutional, time-delay, or Transformer-based encoders. Although ECAPA-TDNN-based models provide strong verification performance, their temporal modeling mainly depends on convolutional and TDNN-style operations. Transformer-based models can capture broader temporal patterns, but they often require high computational and memory costs, making them less suitable for embedded deployment. To address these limitations, this paper proposes HySV, a Hymba-inspired hybrid attention and state-space encoder for deployment-aware speaker verification. Rather than directly employing the original Hymba language model, HySV adapts its hybrid-head principle to speaker embedding extraction. Specifically, conventional ECAPA-TDNN-style encoder blocks are replaced with three stacked Hymba context blocks. Each block contains an attention branch for local speaker-discriminative cue modeling and a state-space branch for efficient temporal context summarization. In addition, a quality-aware decision support module is introduced after cosine similarity scoring to improve reliability using utterance duration, voice activity ratio, and embedding confidence. The proposed system is evaluated using both speaker verification and deployment-oriented metrics, including EER, minDCF, FLOPs, and latency.

키워드