An isolated word recognition system based on DSP and improved dynamic time warping algorithm

Abstract

Dynamic time warping (DTW) algorithm has been widely used in the isolated word recognition system due to its simplicity. But its searching efficiency is low, and it needs large storage space which is inconvenient to be implemented on the digital signal processor(DSP). In this paper, the differential double threshold method is proposed to improve the accuracy of the endpoint detection. The weighted distance searching and vector memory space allocation is proposed to improve the accuracy of the distance calculation and to reduce the storage capacity demand. The matching distance threshold is proposed to stop the matching once the distance of matching is larger than the threshold, thus the redundant matching calculation is cancelled, and the efficiency is improved. The experiments are conducted using the proposed method and 6713DSK board. The results shows that the improved DTW proposed in this paper can obtain a higher recognition rate on the premise of less storage space and faster processing speed compared with the conventional DTW.