This paper presents a system for keyword detection in spontaneous speech. Keywords are predefined through a set of acoustic examples provided by the users. Keyword detection proceeds in two steps: keyword searching and verification. To address the problem of using the same phoneme models in both keyword and filter models, we propose to remove the phoneme models included in the keyword model from the filter models. In order to reduce the false alarms caused by keyword searching step, dynamic time warping (DTW) based template matching and Gaussian Mixture Models (GMM) are proposed. Our keyword detection experiments demonstrate the effectiveness of the proposed methods by yielding improved detection performance compared to the baseline system.