As an important research area in speech recognition, keyword recognition (KR) has broad application space and enormous application value. In this dissertation, the development status and main technologies of KR are demonstrated, and two practical systems are well designed and implemented in application areas of both speech command controlling and audio document content retrieval. The research focuses on the implementation and algorithm comparison of keyword spotting (KWS) system, system design and key technology analysis of these two applications. The main research contents are described in details as follows:1. Research and application of filler based KWS technologyFiller based keyword detecting technology is studied and implemented; comparison is made between different calculation methods in log-likelihood ratio (LLR) based utterance verification; on line garbage score is also used as one of the features to further study the performance of confidence measure. Experiment results show that the refinement of competitive candidate selection can effectively improve the performance of utterance verification, and the combination of both LLR and on line garbage score performs better further.As an application in speech command control area, a filler based speech launch system is implemented, which performs well in experiments. According to project practice, three key problems are well solved: how to modify KWS engine to balance recognition speed and precision while bringing KR technology into practical applications, how to enhance system robust to non-standard mandarin, how to reduce false alarm ratio by engineering method.2. Research and application of syllable confusion network based speech document retrieval technologyThis dissertation describes the development status of audio retrieval technology and 2006 STD testing, realizes a syllable confusion network based KWS engine, which is adopted as recognition core in broadcast audio retrieval system. Inverse Indexing method is used to quickly search in inverted audio document corpus. |