Fig. 3From: An improved i-vector extraction algorithm for speaker verificationA brief figure description to FA-based i-vector adaptation in sparse training data. Different from MAP adaptation in Fig. 2, in the i-vector framework, all components are adapted at the same weight. The adapted mean vectors of upper and lower left components are credible because we indeed have relative speaker frames, but the adaptation of lower right component is questionable because it suffers from a shortage of speaker frames, whose first-order Baum-Welch statistics is inaccurateBack to article page