An integrated pitch tracking algorithm for speech systems
Bruce Gill Secrest, George R. Doddington · 2005
A pitch tracking algorithm is described which operates in the time domain from a conditioned linear prediction residual and applies dynamic programming to optimally determine both pitch and voicing. A set of candidate pitch values are derived from a correlation function applied to an LPC prediction residual which has been low pass filtered in voiced speech and high pass filtered in unvoiced speech by using a single pole filter based on the first reflection coefficient of LPC. A post processing technique using dynamic programming is used to obtain a smooth pitch contour. By incorporating the correlation values of the candidate pitch values, voicing state information and spectral change information into the penalty function of the dynamic programming, a voicing decision is obtained along with an optimum pitch value. This integrated pitch tracking algorithm is compared to three standard pitch tracking algorithms over a data base of 58 male and female speakers ranging from 6 to 87 years of age and is shown to exhibit superior performance.