Discriminative training of GMM parameters for audio scene classification and audio tagging

This report describes the algorithm for audio scene classification and audio tagging and the result for DCASE 2016 challenge data. We propose a discriminative training algorithm to improve the baseline GMM performance. The algorithm updates the baseline GMM parameters by maximizing the margin between classes to improve discriminative performance. For Task1, we use a hierarchical classifier to maximize discriminative performance, and achieve 84% accuracy for given cross validation data. For Task4, we apply binary classifier for each label, and achieve 16.71% EER for given cross validation data.