| Summary: | 碩士 === 國立成功大學 === 電機工程學系碩博士班 === 92 === This study presents two new partial matching algorithms for spoken sentence retrieval and realizes them on PDA and an ARM-based SOPC development board.
For the proposed algorithms, the query and database sentences are initially segmented into equal-size matching units. A matching plane consisting of matching blocks is then created. For each matching block, a local similarity score is then calculated based on the feature distance. The global similarity score of the matching plane indicates the similarity of the query and database sentences. A whole-matching-plane accumulation scheme and a column-based row-based accumulation scheme then are used to obtain the global similarity score. To improve the accuracy of the similarity estimation, the similarity score is calculated through the inverse distance weighting (IDW) technique. The proposed algorithms are based on the feature level comparison and do not require acoustical and language models. The proposed spoken sentence retrieval system thus is language independent. In terms of retrieval performance, the experiments also demonstrate that the proposed spoken sentence retrieval system outperforms the system that uses IBM ViaVoice, a large-vocabulary continuous-speech recognition (LVCSR) system.
After estimating the proposed algorithms, we realize the proposed spoken sentence retrieval system on HP iPAQ Pocket PC and implement the hardware/software co-design version on an ARM-based SOPC development board to be used in various portable speech systems.
|
|---|
