Pre-filtered dynamic time warping for posteriorgram based keyword search

This thesis is motivated by the challange of high computational demand of the dynamic time warping algorithm, which is very prevalent for template based keyword search tasks. The dynamic time warping algorithm suffers from grand calculations of the cost matrix during the search process. To solve this problem, in this thesis, we present a pre-filtering step for the algorithm. We use phonetic posteriorgrams to represent the audio data and generate average posteriorgrams to represent the given text queries. Posteriorgram of the document is divided into segments, or submatrices, and are turned into supervectors, which are used to determine the possible candidates for query matching. Cosine distance is used for the distance measure to find the similarity between the document supervectors and the query supervectors. Filtering is performed according to a threshold value, which specifies the amount of filtering. The other aim is to improve the performance of a large vocabulary continuous speech recognition based keyword search system using the proposed approach for posterior- gram based keyword search. Experimental results indicate that the proposed method reduces the computational complexity of the dynamic time warping algorithm without a significant loss and when combined with the baseline large vocabulary continuous speech recognition based keyword search system, it improves the performance for both the in-vocabulary and the out-of-vocabulary queries.