Fig. 2

A brief figure description of MAP adaptation in sparse training data. As upper Gaussian component has more speaker frames, which corresponds to higher zero-order Baum-Welch statistics N _c , the mean vector of upper component (black dot denotes the mean vector of Gaussian component) will be sufficiently adapted to the normalized first-order Baum-Welch statistics (red cross denotes the adapted mean vector and red dashed ellipse denotes the adapted component). In contrast, sparse speaker frames observed by lower left component correspond to smaller zero-order Baum-Welch statistics and minor adaptation. Lower right component remains still because it has no speaker frame