Variational bayes adapted GMM based models for audio clip classification

The most commonly used method for parameter estimation in the Gaussian mixture models (GMMs) is maximum likelihood (ML). However, it suffers from the overfitting when the model complexity is high. Adapted GMM is an extended version of GMMs and it helps to reduce the overfitting in the model. Variational Bayesian method helps in determining optimal complexity so that it avoids overfitting. In this paper we propose the variational Bayes learning method for training the adapted GMMs. The proposed approach is free from overfitting and singularity problems that arise in the other approaches. This approach is faster in training and allows a fast-scoring technique during testing to reduce the testing time. Studies on the classification of audio clips show that the proposed approach gives a better performance compared to GMMs, adapted GMMs, variational Bayes GMMs. © 2009 Springer-Verlag Berlin Heidelberg.