Trigram duration modeling in speech recognition

Abstract

Rate of speech (ROS) is a very important factor in speech recognition. We present a new speech rate measurement method which first normalizes the duration of different acoustic units to a standard duration and then builds a trigram duration model to measure the speech rate of a sentence. We propose two methods based on the standard duration to compensate the influence introduced by speech rate variation in a data corpus and get 11% error rate reduction in Mandarin digit string recognition.