Detecting user-defined keywords on a resource-constrained edge device is challenging. However, keywords are often limited by a maximum keyword length, which has been largely underexploited in previous work. Our analysis of the keyword length distribution shows that the detection of user-defined keywords can be treated as a limited-length problem, eliminating the need to add variable-length texts. This leads to our proposed method for efficient keyword detection, SLiCK (Subsequence Exploitation for Limited Length Keyword Detection). Additionally, we introduce a subsequence-level matching scheme to learn audio-text relationships at finer granularity, thereby distinguishing similar-sounding keywords more effectively through enhanced context. In SLiCK, the model is trained with a multi-task learning approach using two modules: Matcher (utterance-level matching task, subsequence-level novel matching task) and Encoder (phoneme recognition task). The proposed method improves the baseline results on a hard Libriphrase data set, increasing the AUC from 88.52 to 94.9 and reducing the EER from 18.82 to 11.1.
