Task-Oriented Auxiliary Class Modeling in CTC-Based Speech Keyword Spotting

Speech keyword spotting plays a vital role in the information era with explosive growth of audio data. Hard samples are defined as acoustic segments with feature similarity to keywords that cause misclassification. In this paper, three task-oriented auxiliary classes are proposed to enhance acoustic modeling for: (1) non-keyword segments prone to false alarm, (2) phonetic transition between keyword syllables, and (3) low-quality keyword samples segments. With 0.5 false alarm per hour, the proposed method achieves a relative reduction in false rejection rate (FRR) by 35.9% against Conformer baseline. Ablation study demonstrates that the three auxiliary classes collectively enhance the performance. Robustness analysis demonstrates that the proposed method effectively reduces performance degradation when low-quality keyword samples are present in the training data.