ODYSSEY 2004 - The Speaker and Language Recognition Workshop
May 31 - June 3, 2004
This paper proposes a novel speaker identification method based on the dual Penalized Logistic Regression Machine (dPLRM) for general multi-class discrimination. The machine employs kernel functions which implicitly map an acoustic feature space to a higher dimensional space. Each speaker is discriminatively identified in this space implicitly. The penalized logistic regression model used in dPLRM provides a reliable estimate of probability of each identification decision. Text-independent speech data recorded by 10 male speakers in four sessions over nine months was used to evaluate the new approach. The proposed method effectively reduced the error rate of the conventional GMM-based approach.
Bibliographic reference. Matsui, Tomoko / Tanabe, Kunio (2004): "Speaker identification with dual penalized logistic regression machine", In ODYS-2004, 363-368.