Sequence Classification with Input-Output Hidden Markov Models

We present a training and testing method for Input-Output Hidden Markov Model that is particularly suited for classification of sequences in which class information accumulates over time. We discuss two such cases: the discrimination of mental tasks from sequences of EEG features, common in Brain Computer Interface research, and phoneme classification from sequences of acoustic features for speech recognition. The objective function is modified so that training focuses on the improvement of classification accuracy. For both tasks the algorithm performs significantly better than the alternative solution proposed in the literature, specifically designed for other types of sequences.