Improvement of Speaker Identification by Combining Prosodic Features with Acoustic Features

Abstract

AbstractIn this paper, we study prosodic features derived from pitch parameters to improve the performance of speaker identification (SID) system. In order to deal with the problem of missing pitch in telephone speech, we use pitch estimation for each frame, even in unvoiced regions. After silence frames removal, we also improve prosodic modeling by a weighting form of logarithm of pitch. Then new prosodic features are combined with MFCC parameters. Based on our Gaussian Mixture Model-Universal Background Model (GMM-UBM) recognizer, SID experiments are conducted on the NIST 2001 cellular telephone corpus. Compared to MFCC features, combined features yield 7.0% relative error reduction for male and 2.5% for female. We also discuss the advanced pitch extraction and modeling approach for the improvement of SID systems.KeywordsGaussian Mixture ModelSpeaker RecognitionSpeaker VerificationSpeaker IdentificationProsodic FeatureThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.