New feature extraction methods and the concept of time-warped distance in speech processing

Abstract

For pitch detection, voiced/unvoiced decisions and speech/nonspeech decisions, an improved average magnitude difference function (AMDF) is described that has given promising results: adaptation improves accuracy and skeletonization speeds up computation. A novel definition of time-warped distance results in decreased error probability in speech recognition; however, no fast algorithm for its computation has yet been found. The concept of time-warped average, on the other hand, is easy to compute and results in better speech recognition score. Both improved AMDF and time-warped distance are discussed for use in the speaker identification environment.<>