Designing an ultra-hispeed search engine for big data of spoken documents

点击翻译按钮获取中文摘要

Fast spoken term detection from big data has been developed. In the development, we have focused on (1) accurate speech-to-phoneme conversion even though there is an out-of-vocabulary word in an utterance, and (2) fast spoken term detection even though there is ambiguity caused by speech recognition errors.In (1), after extracting features of phonemes in a dual space, a multi-layer perceptron extracts articulatory features, then a subsequent phoneme classifier discriminates phonemes with high accuracy. In (2), we have implemented an iterative deepening search based on suffix arrays, a continuous DP matching, and a keyword division algorithm. As a result, we could solve three issues of search accuracy, search speed, and size of index at the same time.