Classification of varying length multivariate time series using Gaussian mixture models and support vector machines

In this paper, we propose two approaches in a hybrid framework in which a Gaussian mixture model (GMM) based method is used to obtain a fixed length pattern representation for a varying length time series and then a discriminative model is used for classification. In score vector based approach, each time series in a training data set is modelled by a GMM. A log-likelihood score vector representation is obtained by applying a time series to all GMMs. In segment modelling based approach, a time series is segmented into fixed number of segments and a GMM is built for each segment. Parameters of GMMs of segments are concatenated to obtain a parametric vector representation. Support vector machine is used for classification of score vector representation and parametric vector representation of time series. The proposed approaches are studied for speech emotion recognition and audio clip classification tasks.