Score normalization for keyword search
Leda Sarı, Murat Saraçlar · 2016
In this work, keyword search (KWS) is based on a symbolic index that uses posteriorgram representation of the speech data. For each query, sum-to-one normalization or keyword specific thresholding is applied to the search results. The effect of these methods on the proposed KWS system is investigated. Results are combined with a KWS system that is based on an index generated from automatic speech recognition (ASR) lattices and the effect of score normalization on the performance of the combined system is observed. Maximum term weighted value (MTWV) is used as the performance measure. It is shown that the MTWV is increased in the combined system especially for the out-of-vocabulary (OOV) keywords that are not covered by the ASR lexicon.