A new word pre-selection method based on an extended redundant hash addressing for continuous speech recognition

This method is based on the isolated word recognition method using the redundant hash addressing principle proposed by T. Kohonen et al. (Inf. Sci. vol.33, p.3-30 of 1984). The redundant hash addressing method cannot be used for continuous speech recognition because the word boundary in continuous speech is not known. Kohonen´s method has been extended to continuous speech recognition. The proposed method spots words from the input speech using information about locations of features in reference patterns. The combination of this method and an ordinary parser makes a speech recognition system faster. Experimental results are shown, comparing this method with the continuous-DP matching method, a traditional word spotting method. While the proposed method gives performance comparable with that of continuous-DP matching, it is five times faster.<>