Inter-Class MLLR for Speaker Adaptation

In transformation-based adaptation, increasing the number of transformation classes can provide more detailed information for adaptation, but at the expense of greater estimation error with small amounts of data. In this paper we introduce a new procedure, inter-class MLLR, which utilizes relationships among different classes to achieve more reliable estimates of the transformation parameters across multiple classes using limited adaptation data. In this method, the inter-class relation is given by a linear regression which is estimated from training data. In experiments using non-native English speakers from the Spoke 3 data in the 1994 DARPA Wall Street Journal evaluation, interclass MLLR provided a relative reduction in word error rates of 15.3 % compared to conventional MLLR with a small amount of adaptation data. TRANSFORMATION-BASED ADAPTATION ◆ Conventional transformation-based adaptation: • More effective than Bayesian approaches for small amounts of adaptation data • Each transformation class is considered independently • With limited adaptation data we cannot simultaneously achieve both • highly reliable estimates of the transformation parameters A and b • estimates of the Gaussian mean vectors that can be different for each transformation class ◆ Goal of this work: • Utilize relationships among transformation classes to achieve more reliable estimates of the transformation parameters across multiple classes Single class All data available for reliable estimation of transformation function, but all Gaussians must move in the same direction Multiple classes (independent) Each class can move in a different direction, but less data are available within each adaptation class Multiple classes (inter-related