This work describes a new partial matching algorithm to retrieve speech information from a speech database by spoken sentence query. The similarity between two partial matching sentences is evaluated by a new algorithm called Column-Based Row-Based (CBRB) evaluation. This feature-matching method does not need a complex language model, so the system is language independent. Moreover, the proposed approach does not involve a large-vocabulary recognizer, which can heavily load a system, so the proposed approach greatly reduces the computational burden, and is highly appropriate for handheld devices. Finally, experiments were conducted for personal speech calendar and the results show our system that can successfully evaluate the similarity between database sentences and query sentence.
