A Robust Keyword Spotting System for Mandarin Speech

This paper introduces a method for designing a robust Mandarin keyword spotting system. Keywords which will be extracted from an uttered sentence are modeled by sequences of states. These state models that represent the subsyllables of Mandarin speech are generated by using the existing speech database. The non-keyword portions of an input utterance are filtered out by filler models. A simplified signal bias removal technique is applied to overcome the influences due to channel distortion and speaker variation. State integrated Wiener filters are used for noise compensation. Proposed techniques are evaluated by several experiments to show their effectiveness for robust speech recognition.