On the Initialization of Dynamic Models for Speech Features

Abstract:
In this work, a novel approach for the initialization of switching linear dynamic models (SLDMs) as dynamic models for the trajectory of speech features is proposed. Borrowing ideas from the "k-means++"-algorithm, the goal of this approach is to find distinctly different SLDMs, modelling the complex dynamics of the speech features, already at the initialization stage of a subsequently following "expectation-maximization (EM)"-algorithm. Experimental results comparing differently initialized SLDMs in a model-based speech feature enhancement scheme show the superiority of the proposed initialization routine in terms of a reduced word error rate on an automatic speech recognition task.