Hybrid word-subword spoken term detection

The thesis investigates into keyword spotting and spoken term detection (STD), that are considered as sub-sets of spoken document retrieval. It deals with two-phase approaches where speech is first processed by speech recognizer, and the search for queries is performed in the output of this recognizer. Standard large vocabulary continuous speech recognizer (LVCSR) with fixed vocabulary is not capable incapability of detecting out-of-vocabulary words (OOV). A hybrid spoken term detection system combining both word and subword parts in one recognition network is proposed. Extensive experiments investigating into different variants of this approach are performed, and the results (in terms of spoken term detection precision, speed, and necessary computing resources) are reported on standard data from NIST STD 2006 evaluation.