Open set text-independent speaker recognition based on set-score pattern classification

We propose a two-stage recognition schema for open set text-independent speaker recognition tasks. First we try to find a best matched model (which gets the best score) for the unknown speaker like many other systems. But then unlike other classical threshold selecting methods that make decisions based on the best score, we use the scores over a reference speakers set as a whole (called the set-score pattern): a binary classifier (e.g., an SVM) is then built to recognize acceptable and rejectable patterns. The results show that the set-score pattern classification method gives reasonably good performance. An obvious improvement has been seen compared to simple threshold selecting methods. And the painful procedure to choose a good threshold can be avoided too.