Model Based Clustering of Audio Clips Using Gaussian Mixture Models

Abstract

The task of clustering multi-variate trajectory data of varying length exists in various domains. Model-based methods are capable of handling varying length trajectories without changing the length or structure. Hidden Markov models (HMMs) are widely used for trajectory data modeling. However, HMMs are not suitable for trajectories of long duration. In this paper, we propose a similarity based representation for multi-variate, varying length trajectories of long duration using Gaussian mixture models. Each trajectory is modeled by a Gaussian mixture model (GMM). The log-likelihood of a trajectory for a given GMM model is used as a similarity score. The scores corresponding to all the trajectories in the given data set and all the GMMs are used to form a score matrix that is used in a clustering algorithm. The proposed model based clustering method is applied on the audio clips which are multi-variate trajectories of varying length and long duration. The performance of the proposed method is much better than the method that uses a fixed length representation for an audio clip based on the perceptual features.