Abstract:

This talk will start with a short presentation of IDIAP, its research groups (speech recognition, computer vision, and machine learning), and a short description of some of the current projects.

The second part of the talk will mainly deal with an overview of speaker verification, including recent developments at IDIAP towards text dependent speaker verification using decision tree based binary classifiers. More specifically, we will show how a speaker verification task can be advantageously decomposed into a large number of binary classification problems (speaker/anti-speaker models), each decision tree (dealing with attributes of continuous values) being specific to one speaker and one word. The resulting set of decision trees is then pruned according to a simple criteria (maximizing the discrimination between speakers) to eliminate the less relevant trees. It will be shown on NIST evaluation data that the resulting system is at least as good as state-of-the-art speaker verification systems, while being more robust to a priori setting of the rejection threshold.