Robust Biometric Identification Combining Face and Speech

This paper deals with two problems: the problem of combining different classification systems in order to obtain a combined system that outperforms each one alone and the problem of combining classification results from different sources (i.e. image and speech). In the first problem the classification systems must be based on different philosophies (i.e. VQ, projection matrices, neural net, etc.), in order to have different susbsets of mistakes, and use a decision module that yields a performance of the whole system equal to the intersection of mistakes of each classifier alone. In the second problem the classification system must use the fact that different sources of information (image or speech) will probably have different performance on different individuals, that is, classify each individual with different certainties. Also robustness will be added, because each information will be affected by different sources of noise, and one source can compensate the degradation of the other.