Real-time end point detection specialized for acceleration signal

Due to temporal and spectral difference between speech and acceleration signal, the conventional end point detection (EPD) in automatic speech recognition cannot be directly applied to acceleration and threshold-based algorithms found in literatures are too heuristic to be accepted for automatic EPD. In this regard, for motion detection by acceleration, supervised learning in pattern recognition is proposed to discriminate a motion state and a non-motion state simply. In succession of the previous research where we´ve concentrated on the feasibility test of the proposed approach and feature selection in general pattern recognition procedure, a new recognizer, radial basis function network (RBF), is subsequently designed for the performance comparison with multi-layer perceptron (MLP) which serves as a performance baseline. As a result, it is reported that the recognition rates variance between feature vectors is not significant in RBF while it is significant in MLP. In addition, recognition rates variance between subjects shows clear difference statistically in the both ways but more serious in RBF. Finally it is concluded that MLP and RBF don´t make significant recognition rates difference and confirmed again that the sequence of the absolute 1st derivatives record comparatively more reliable and stable recognition performance.