Analysis and improvement of auto-correlation pitch extraction algorithm based on candidate set

Abstract

The auto-correlation pitch extraction algorithm based on candidate set is widely adopted due to low complexity, strong noise immunity and high accuracy. However, there are still some discontinuous or unsmooth places in the pitch contours got by this algorithm. This paper analyses types and causes of error pitches in the results of such algorithm, and proposes detection methods and amendment strategies for each type of error. Using the pitch contour smoothing algorithm based on segmenting, 91.07 percent of error pitches in the result are eliminated. © 2012 Springer-Verlag Berlin Heidelberg.