Comparing decoding strategies for subword-based keyword spotting in low-resourced languages

International audience For languages with limited training resources, out-of-vocabulary (OOV) words are a significant problem, both fortranscription and keyword spotting. This paper investigates theuse of subword lexical units for keyword spotting. Three strate-gies for using the sub-word units are explored: 1) convertingword-based lattices to subword lattices after decoding, 2) per-forming a separate decoding for each subword type, and 3) asingle decoding using all possible subword units. In these ex-periments, the best performance is achieved by carrying out aseparate decoding for each subword type. Further gains are at-tained through system combination. We also find that ignor-ing word boundaries improves the detection of OOV keywordswithout significantly impacting in-vocabulary keyword detec-tion. Results are presented on four languages from the IARPABabel Program (Haitian Creole, Assamese, Bengali, and Zulu).