INTERSPEECH 2004 - ICSLP
This paper investigates a probabilistic speaker identification method based on the dual Penalized Logistic Regression Machines (dPLRMs). The machines employ kernel functions which map an acoustic feature space to a higher dimensional space as is the case with the Support Vector Machines (SVMs). Nonlinearity in discriminating each speaker is implicitly handled in this space. While SVMs maximize the margin between two classes of data, dPLRMs maximize a penalized likelihood of a logistic regression model for multi-class discrimination. dPLRMs provide a probability estimate of each identification decision. We show that the performance of dPLRMs is competitive with that of SVMs through text-independent speaker identification experiments in which speech data recorded by 10 male speakers in four sessions are analized.
Bibliographic reference. Matsui, Tomoko / Tanabe, Kunio (2004): "Probabilistic speaker identification with dual penalized logistic regression machine", In INTERSPEECH-2004, 1797-1800.