Proc. SPIE. 5813, Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications 2005
KEYWORDS: Information fusion, Biometrics, Data modeling, Data hiding, Feature extraction, System identification, Acoustics, Performance modeling, Speaker recognition, Expectation maximization algorithms
Prior knowledge helps to make the speaker recognition system more reliable and robust. This paper presents a uniform framework of feature-level fusion to incorporate the prior knowledge for speaker recognition using gender information based on dynamic Bayesian network (DBN). DBNs are a new statistical approach, with the ability to handle hidden variables and missing data in a principled way with high extensibility. And thus, DBNs can describe the prior knowledge conveniently. Our contribution is to apply DBNs to construct a general feature-level fusion to combine the general acoustic feature like MFCC and prior information like gender into a single DBN for speaker identification. In our framework, gender information become additional observed data to influence both hidden variables and observed acoustic data. Experimental evaluation over a subnet of YOHO corpus show promising results.