Keyword Spotting using Keyword Adapted Language Model

Abstract

In keyword spotting applications, language modeling directly affects the system performance, as well as the acoustical modeling. This study focuses on the effects of different language models on the keyword spotting performance on Turkish voice recordings. Three different systems, one of which is proposed by us, that use different language models are tested and compared. Two of the tested systems are the conventional HMM based systems, one using a language model, and the other one using a no-grammar language model. The third system, which we propose, selects the sentences with the searched keywords from a big text database, and trains the language model with this new keyword-specific text, with the most frequently occurred words in it. Tests showed that, the system that uses keyword adapted language model gives the best performance in both recall and detection time. Highest recall rate is 86%. The system we proposed has increased the recall performance by 4% from the system that uses language model, and by 13% from the system that uses no language model. However the system that we propose, gives lower precision results than the other two systems, but precision can be increased by some language model processing techniques like smoothing or interpolation.