Spotting recognition is the simultaneous realization of both recognition and segmentation. It is able to extract suitable information from an input dataset satisfying a query, and has developed into a research topic known as word spotting that uses dynamic programming or hidden Markov models. Continuous dynamic programming (CDP) is a promising method for spotting recognition applied to sequential patterns. However, the computational burden for conducting a retrieval task using CDP increases as O(JIP), where I is the input length, J is the reference length and P is the number of paths. This paper proposes a faster nonlinear spotting method like CDP, called Fast Spotter (FS). FS is regarded as an approximation of CDP using A{sup}* search. FS reduces the computational burden to O(IP log_2 J) in the best case and executes in around half the time with an experimental dataset, enabling it to realize a large-scale speech retrieval system.
展开▼
