Modeling dependency between regression classes in MLLR using multiscale autoregressive models

Adapting acoustic models to a new environment is usually realized by considering model transformations that are estimated on the adaptation corpus. Since such a corpus usually contains very few data, the models' Gaussians are most often partitioned into a few regression classes, and all the Gaussians in the same class share the same transformation. It is further possible to increase the number of transformations by modeling the dependency between the regression classes. We present, in this paper, such a technique where dependency is modeled by multiscale autoregressive (MAR) processes. The power of the MAR framework resides in its ability to efficiently and optimally estimate the state vector at each node of the regression tree, based on sparse and noisy measurements at different resolutions. The method is evaluated on a french numbers recognition task where the test corpus has been recorded in a car at various speeds and noise levels. The proposed adaptation method is based on Maximum Likelihood Linear Regression.