Crowd-supervised Speaker Identification

Training state-of-the-art machine learning methods for the task
speaker identification typically requires a large annotated corpus of
labeled speech. In this research we are developing methods that
require as few labels as possible to obtain the same level of
performance as a fully supervised setup. For this UROP project, we
are looking to explore the use of crowdsourced-based labeling as a
realistic means to annotate data. Over the summer, this UROP will
involve developing human intelligence tasks that can be deployed on
Amazon Mechanical Turk, and subsequently incorporated into an
active-learning-based method for speaker recognition. Interested
students should send a CV to Jim Glass,