Depitch and the role of fundamental frequency in speaker recognition

摘要: Pitch information is known to be partially conveyed in Mel cepstral features that are commonly used for speaker recognition. In particular, high pitched female speakers, and whenever average pitch varies significantly between enrollment testing, the fine spectral structure introduced by fundamental frequency was shown degrade recognition performance. This paper introduces a signal processing procedure termed depitch attempts remove from speech signal. Recognition experiments carried out on subset of NIST 2002 Speaker Evaluation show combining scores conventional depitched system, substantial improvement equal error rate obtained speakers pitch-mismatched trials. Performing pitch/depitch score fusion also help alleviate well-known problem "goat" speakers.