Connected Spoken Digit Recognition by Augmented Continuous DP Matching and its Evaluation.

Abstract

Recently, we proposed the Augmented Continuous dynamic time warping algorithm for connected spoken word recognition. The algorithm is based on the same principle as the Two Level DP and Level Building DP. Our algorithm obtains a near optimal solution for the recognition principle based on pattern matching. However, it is computationally more efficient than the conventional methods and does not require much memory storages. In this paper, we modified this algorithm on the following points : 1) matching methods for a silence in speech, 2) scoring mechanisms for connected words, 3) simple word spotting algorithms, 4) generation methods of the best plural candidates and 5) endpoint-free matching mechanism. We apply this modified algorithm to connected spoken digit recognition and evaluate it. We find from experiments as follows : First, the consideration of speech silence improves the recognition performance. Secondly, the matching result of longer reference is more important than that of shorter reference in the unknown case of number of words in a test sample. Thirdly, the recognition results by simple algorithms for detecting two word spotting candidates are comparable with the results by the original complex algorithm. Forthly, the optimal strategy for the generation of the best plural candidates of word sequences is better than the suboptimal strategy. Fifthly, the endpoint-free matching mechanism improved the recognition rate. By using one reference pattern per digit, we obtain the string recognition rate of 89.0% with the top choice and 95.3% with the top 3 choices in a speaker-dependent mode. Finally, we compare this algorithm with traditional algorithms such as the Two Level DP and find that this algorithm has almost the same performance as the traditional algorithms.