Research Of Problems In Spoken Term Detection

With the increasingly mature of speech recognition and explosive growth of multimedia data, the speech retrieval technology get more and more people’s attention and widespread applied. There are two difficult problems:speech recognition error and out-of-vocabulary problem, the two problems influence the precision and recall of speech retrieval. For the two problems, this paper focuses on index structure, query expansion, semantic analysis to develop the performance of spoken term detection. Its main contributions and innovations are described as follows:1. The form and application of two-layer index retrieval structureThis paper transform the word confusion to syllable confusion network, then create a two-layer index based on one-best and syllable confusion network. The experiments show that two-layer index structure can develop the recall rate with the precision stable, generated syllable confusion network brings in the limitation of words, and develop the compatibility of homonyms, and the information of one-best-layer index can develop the performance of detection.2. Query expansion based on confusion matrixConfusion matrix trained through one-best recognition result and word confusion network is applied in input to query expansion, and brings in syllable-term frequency, the experiments show that based on the index, the bring-in of query expansion with syllable model can recall some recognition error in some ratio, and develop the retrieval precision compared to the normal query expansion.3. Retrieval filter algorithm based on language model and word activation forceWith the two-layer index structure, the bring-in of language model and word activation force model which denote long distance information can filter some fake alarm, experiments show that language model and word activation force can sort and filter retrieval results to develop the performance.