ODYSSEY 2004 - The Speaker and Language Recognition Workshop
May 31 - June 3, 2004
We will treat a task of statistical detection using discrete alphabet in this paper. It is well known that the likelihood ratio detector is the optimal one, provided that "measurements" taken on the detected object are mutually independent and that we know the feature distributions of target and background objects precisely. The detector presented here is optimal using only the independence assumption. Instead of requiring the knowledge of the underlying distributions it relies on the training data itself. Further, we introduce the averaging technique which aims to lower the effects of statistical dependence. This averaged detector outperformed similarly averaged likelihood ratio detector by 7% relative in the task of speaker detection.
Bibliographic reference. Klusácek, David (2004): "Optimal detection in case of the sparse training data", In ODYS-2004, 97-104.