Spiking Neural Network to Extract Frequent Words from Japanese Speech Data

Abstract This study aimed to automatically extract frequent words from speech data in Japanese. The length of words that can be extracted by the previous method was up to 2 symbols length. So, in this paper, we aimed to extract sub-sequences longer than 3 symbols. To extract sub-sequences longer than 3 symbols length, we proposed a new structure based on the neural network of the previous method. The new structure can extract longer sub-sequences by repeatedly stacking the structure which extracts sub-sequences of two symbol length. In order to confirm that the proposed method can extract frequent words from speech data by real time processing, we gave reading aloud data of Japanese to the proposed neural network and confirmed that the neural network can extracted frequent words. We confirmed that this neural network extracted a frequent word of 4 symbol length.