Using formant frequencies to word detection in recorded speech

The paper considers increasing the precision of detection of words in unsupervised keyword spotting method. The method is based on examining signal similarity of two analyzed media description: registered voice and a word (textual query) synthesized by using Text-to-Speech tools. The descriptions of media were given by a sequence of Mel-Frequency Cepstral Coefficients or Human-Factor Cepstral Coefficients. Dynamic Time Warping algorithm has been applied to provide time alignment of the given media descriptions. The detection involved classification method based on cost function, calculated upon signal similarity and alignment path. Potential false matches were eliminated in the algorithm by applying two-staged verification, using the Longest Common Subsequence algorithm and analyzing formant frequencies of eleven English monophthons. The use of formant frequencies at the stage of verification increased overall detection precision by about 10% as compared to original algorithm.