christegho/keywordspotting

Keyword spotting (KWS) with written queries on low resource languages # keywordspotting This practical investigates keyword spotting (KWS) with written queries on low resource languages. The Swahili corpus from the IARPA Babel project's 2015 surprise evaluation language is used. A KWS system is implemented and 1-best ASR decoding output from a word-based system and a morph-based system are used. Mapping out-of- vocabulary keywords to proxy keywords using graphemic confusion networks and the use of score normalization are investigated. Outputs from a WFST-based KWS system are also used and their hits and scores combined with the systems using 1-best output. The eect of score normalization on system combination is also investigated. Complete Report: github.com).pdf