Bidirectional Alignment of Glottal Pulse Length Sequences for the Evaluation of Pitch Detection Algorithms

摘要

This paper describes a problem in a reported Dynamic Time Warping (DTW) alignment procedure to compare the reference and detected glottal pulse length sequences, oriented to compare the evaluation of Pitch Detection Algorithms (PDAs) in pathological voices. The problem in the existing alignment method tends to overestimate the failure of the PDA, by aligning only the detected to the reference sequence. A solution is presented, which performs a bidirectional alignment reducing the differences present in the definitive comparison. The proposal is evaluated in both synthetic and real voice signals, by running three well-known PDAs, and the magnitude of the error reduction along with comments on the possible factors influencing its value, are given. The alignment variant introduced in this paper allows to perform a fairer comparison of the PDAs performances.