Multi-environment Speaker Verification

Here we investigate an instance of the abstract problem of pattern recognition under mismatch conditions: Models of phenomena are built with data collected in the training environment but must be used to recognize the same phenomena in another environment. The speciic problem is speaker veriication, where the training and testing data for each speaker can come from one of many diierent microphones. We use data, unlabeled with respect to channel or environment, to build, unsupervised, an easily extensi-ble, hierarchical structure that at the nest level consists of individual speaker models, but at the coarsest level is a collection of all of the models. We then have the ability to automatically generate evolving background models from any layer of our hierarchical model when we wish to perform a veriication. We give results to show that the richer our hierarchical structure, the better we do in terms of veriication. 1. INTRODUCTION We consider the problem of speaker veriication under mis-matched conditions when the number of environments in which training and testing data can be collected is large and variable. We describe our technique for dealing with multiple environments, making a special note of the fact that it is unsupervised and incremental, and so can potentially be used in situations where we do not know beforehand, the nature of the environments in which we will be collecting data. For example, take situations in which the only way we know about new environments is through enrollment data. In the multi-environment veriication context, the technique is best described as one allowing easy modiication of the normalizing background to reeect data from perhaps new and unknown environments, as those environments are seen in enrollment. Speciically, we analyze performance on data collected form 8 diierent microphones in a relatively noisy environment, a cafeteria. We show that we are able to obtain an improvement in multi-environment speaker recognition performance by adding information about multiple environments solely through our enrollment process, which is eecient. Traditional approaches to such normalization in speaker veriication have involved the supervised use of data from