OPC-KWS: Optimizing Keyword Spotting with Path Retrieval Decoding and Contrastive Learning

As voice interaction capabilities with smart devices advance and the demand for personalized wake words increases, customized keyword spotting (KWS) has become critically important. Recent models for customized keyword spotting have effectively utilized RNN-T or attention mechanisms, introducing unique decoding strategies for device activation. However, these methods typically involve high parameter counts and complex workflows. In response, our study introduces a lightweight framework, OPC-KWS, which aims to achieve similar performance using efficient, small-scale models and simple training methods. Our method employs contrastive learning to achieve more robust sound representations. Additionally, by redesigning the traditional algorithm for decoding, we have developed an effective approach to detect keyword sequences. Our proposed method can achieve similar or even better performance with only 29% of the parameters, achieving a favorable False Acceptance Rate (FAR) of 0.07.