Fast Keyword Spotting in Telephone Speech

Abstract

In the paper, we present a system designed for detecting keywords in telephone speech. We focus not only on achieving high accuracy but also on very short processing time. The keyword spotting system can run in three modes: a) an off-line mode requiring less than 0.1xRT, b) an on-line mode with minimum (2 s) latency, and c) a repeated spotting mode, in which pre-computed values allow for additional acceleration. Its performance is evaluated on recordings of Czech spontaneous telephone speech using rather large and complex keyword lists.