WEakly supervised hmm learning for spokenword acquisition in human computer interaction with little manual effort

In this paper, weakly supervised HMM learning is applied to modeling word acquisition towards human-computer interaction with little manual effort. The only imposed supervisory information is initializing the learning algorithms by two labeled data samples per pattern. Experiments on TIDIG-ITS show that our recently proposed algorithm, Baum-Welch learning regularized by non-negative Tucker decomposition, succeeds in finding good solutions in the sense of yielding high recognition accuracy on the testing data which approximate the supervised baseline (98.0% vs 98.9%).