The aim of this work is to improve the performance of an existing KWS system by merging the search results produced by two additional KWS systems. The existing baseline system is based on large vocabulary continuous speech recognition (LVCSR) and weighted finite state transducers (WFST). The first proposed KWS system is based on searching a symbolic WFST index which is generated by quantizing the posteriorgram representation of the audio. The second proposed KWS system is based on subsequence dynamic time warping (sDTW) algorithm which is commonly used in the query-by-example spoken term detection (QbE-STD) tasks. We also investigate using average posteriorgrams for query generation. Experimental results show that when combined with the existing KWS system, the proposed systems improve the performance of the KWS system especially for the out-of-vocabulary (OOV) queries.
Bibliographic reference. Sarı, Leda / Gündoğdu, Batuhan / Saraçlar, Murat (2015): "Fusion of LVCSR and posteriorgram based keyword search", In INTERSPEECH-2015, 824-828.