Exploring Fusion Methods and Feature Space for the Classification of Paralinguistic Information

This paper introduces the different systems developed by Aholab Signal Processing Laboratory for The INTERSPEECH 2017 Computational Paralinguistics Challenge, which includes three different subtasks: Addressee, Cold and Snoring classification. Several classification strategies and features related with the spectrum, prosody and phase have been tested separately and further combined by using different fusion techniques, such as early fusion by means of multi-feature vectors, late fusion of the standalone classifier scores and label fusion via weighted voting. The obtained results show that the applied fusion methods improve the performance of the standalone detectors and provide systems capable of outperforming the baseline systems in terms of UAR.