Speech recognition system, speech recognition method, and speech recognition program

Abstract

Speech recognition is performed at high speed by processing respective paths of multi-path speech recognition in parallel. A distance calculation unit 111 receives temporal data sequence of acoustic features, calculates the distance values between all acoustic models and the speech features in the respective frames, and writes them in a distance value buffer 141. An acoustic lookahead unit 121 receives distance values from a plurality of distance value buffers 144 to 142, calculates lookahead values which are relative priorities of respective recognition units, and writes them into the lookahead value buffer 142. A word string matching unit 131 receives information from a plurality of distance value buffers 145 to 143 and the lookahead value buffers 153 to 152, and recognizes the entire utterance in frame synchronization by adequately selecting matching words using the lookahead values to thereby generate a recognition result.

Description

technical field [0001] The present invention relates to a speech recognition system, a speech recognition method, and a speech recognition program suitable for, for example, high-accuracy and high-speed Large Vocabulary Continuous Speech Recognition (LVCSR). Background technique [0002] As described in Non-Patent Document 1, in recent years, attempts to realize high-accuracy and high-speed LVCSR have become increasingly active. In this LVCSR, since the search space is very large, the design of the search algorithm is important. In this search algorithm, a method of "predicting" not only the cumulative score of a node on a trellis but also the cumulative score (estimated value) after the node is considered during the search is widely used. A speech recognition system related to this technique is shown in FIG. 5 as a first conventional example. Hereinafter, it demonstrates based on this drawing. [0003] The speech recognition system 700 of the first conventional example i...

AI Technical Summary

Problems solved by technology

Therefore, the technique of the second conventional example cannot be directly combined with the first conventional example including such prediction processing that needs to store a plurality of frames in a buffer and then process them in reverse order of time.

Images