Few-Shot Open-Set Keyword Spotting with Multi-Stage Training

As the advance of human-computer interaction technologies continued, keyword spotting (KWS) systems have gained prominence in everyday devices. This study is dedicated to exploring innovative approaches for few-shot keyword recognition under open-set conditions, a challenging yet crucial area in speech processing. To this end, we design and develop a multi-stage training method that synergistically combines the advantages of acoustic and phonetic features, thereby substantially enhancing the ability of a KWS model. By learning multi-type features with joint training from only one dataset, our KWS model is equipped with a more robustness feature extractor to deal with few-shot KWS. Experimental results demonstrate that our model outperforms strong baselines by achieving a 15% improvement in recognition accuracy on open-set tests in a 10shot-10way setting. This research confirms the effectiveness of our multi-stage strategy and suggests promising directions for future development in keyword recognition technologies.