Structured Approaches to Data Selection for Speaker Recognition
Howard Hao Lei and Nelson Morgan
International Computer Science Institute
In text-dependent speaker recognition, the aim is to perform speaker recognition by only using portions of speech signals where certain speech units (i.e., phone N-grams, word N-grams) exist. In this work, we investigate various measures to determine which ones correlate well with speaker recognition performance of certain speech units. The measures will be used to perform arbitrary data/unit selection for speaker recognition, where the data/units will be selected based on their speaker discriminative abilities. Various issues will be investigated, such as the types of features used, speaker gender differences, coping with missing speech units, the robustness of certain speech units to undesired variability, and unit combination. This project can also lead to better prediction of speaker recognition performance based on data usage.