The importance of optimal parameter setting for pitch extraction.

Abstract

In this study we present a performance comparison for five pitch extraction algorithms: auto correlation (AC), cross correlation (CC), and sub-harmonic summation (as implemented in PRAAT [Boersma and Weenick (2010)]), the robust algorithm for pitch tracking (RAPT) implemented in ESPS [Talkin (1995)], and SWIPE’ [Camacho (2007)]. Recent research showed that SHS and SWIPE’ outperformed the other algorithms on two speech databases with EGG reference values [Camacho (2007)]. That study, however, used a fixed search range of 40–800 Hz for all speakers, regardless of sex or speaker-specific pitch characteristics. In the current study, we adopt the parameter optimization strategy from De Looze and Rauzy (2009) to calculate specific pitch floor and ceiling values for each speaker. Our results show a substantial improvement in accuracy of the AC, CC, and RAPT algorithms when the optimized parameters are used (especially for the female speakers), and all five algorithms show similar performance. The gross error rate for all five algorithms ranges from 0.1% to 0.3% (N=18 098) on the FDA database [Bagshaw (1994)] and from 0.2% to 0.4% (N=11 527) on the KPD database [Plante et al. (1995)]. Our study thus highlights the importance of pre-processing the speech signal to determine optimal speaker-specific parameters for pitch extraction.