A Hokkien isolated word recognition system has been implemented on the Spartan-6 XC6SLX45 FPGA hard-core in this paper. Firstly, an efficient method using adaptive double threshold is employed to execute the voice available detection. Secondly, the 24-th order cepstral analysis of static MFCC and differential MFCC is used to extract speech signal features. Finally Dynamic Time Warping (DTW) is utilized to find the minimum cost path between test and reference templates. The recognition system works offline and is built on medium class FPGA. The experiments on specific and non-specific speaker recognition for Hokkien isolated digits were implemented. And the analyzed results showed that the proposed system achieved a good performance in terms of recognition accuracy, calculation speed and hardware utilization.

No abstract

No abstract

In this paper, a novel approach is proposed for shape classification based on the semi-supervised framework. For shape similarity-measuring problem, in order to avoid solving an NP problem produced by finding some affine transformation and to enhance its robustness for local changes of the shapes, we switch to compute an energy index defined by the degree of segmentation. The corresponding segmentation is achieved by active contour modeling with the low-rank constraint by the prior shape. Then, the sequence of shapes is classified into a certain number of categories by repeating this scheme in a semisupervised framework. The experiment results showed the feasibility of our model.