Mirex 2015: Methods for Speech / Music Detection and Classification

With this submission, a set of ensemble learning based methods for the MIREX 2015 Speech / Music Classification and Detection task is proposed and evaluated. The main algorithm for the Detection task employs a self similarity matrix analysis technique to detect homogeneous segments of audio that can be subsequently classified as music or speech by a Random Forest classifier. In addition to the main algorithm two variations are proposed, the first one employs a silence detection algorithm while the second one omits the self-similarity information and relies solely on the Random Forest classifier. For the Classification task two variants are proposed, both based on a sliding-window classification approach. In the first case a pre-trained model is used, while in the second case, a training phase exploiting training data provided during the submission evaluation, precedes classification.