96 A Comparison of Thai Keyword Spotting Systems Using Different Types of Features

Two methods for Thai keyword spotting were studied. The first method was based on a HMM-based LVCSR system and the second was the discriminative keyword spotting technique which based on large margin and kernel method. Each method was evaluated on a Thai keyword spotting task using 19 different feature sets. LOTUS speech corpus was used to conduct all experiments. The objective of this work was to evaluate the performance of both keyword spotting methods on each type of feature set. Experimental results showed that the discriminative keyword spotting approach that utilized DCTC/DCS features outperformed the LVCSR system (⁠AUC = 0.57 versus 0.51). One major drawback of the discriminative keyword spotting method was its computation inefficiency.