Audio to the Rescue

Abstract

Automatic recognition of people is a challenging problem which has received much attention during the recent years due to its many applications in different fields such as law enforcement, security applications or video indexing. Recognition of people can be achieved using different biometrics such as face, voice, fingerprints or iris scans among others. In most cases, the choice of the particular biometric deeply relies on the final application. For instance, retinal scans have shown high recognition accuracy, however their use is limited to the availability of cooperative individuals, which is not always possible. Although relatively high recognition rates have been obtained using separated biometrics, there are some limitations that make very difficult to increase the recognition performance using individual modalities. Examples of these limitations are changes in illumination or pose for the face biometric, and ambient noise and channel distortion for the voice biometric. Although, more powerful recognition schemes for each modality would probably improve the recognition rates, there is another potential way to increase the recognition performance which consists in combining recognition results from different biometrics. The key idea behind this combined approach is that different information sources can complement each other since degradations for each modality are usually uncorrelated. A good example of a system that combine multiple information sources is the human being, e.g. it has been shown that simultaneously seeing and listening a person talking greatly increases intelligibility. In this paper, we will focus on the recognition of people in video sequences for video indexing applications. This is: given a video sequence, we want to locate those clips where a particular person