Keyword spotting in continuous speech utterances

The work in this thesis constructed a word spotting system, which managed to spot an amount of pre-defined keywords out of unconstrained running conversational speech utterances. The development and experiments are based on the Credit Card subset of SWITCHBOARD speech corpus. The techniques are applied in the context of a Hidden Markov Model (HMM) based Continuous Speech Recognition (CSR) approach to keyword spotting. The word spotting system uses context-dependent acoustic triphone to model both keyword and non-keyword speech utterances. To enhance the true keyword spotting rate, sophisticated keyword-filler network topology models are defined in two different orthographic ways, individual phonemic filler models and individual syllabic filler models. To introduce more lexical constraints, a bigram language model is used. Better performance is obtained in the system with more lexical constraints. A background acoustic model is paralleled to the system network to account for the acoustic variety. The results of the experiments show that the word spotting rate of the overall performance increased by 84% when more lexical constraints applied, and the merge of the background model helps to increase the spotting rate by 5.73%.