Analysis of Deep Learning-Based Decision-Making in an Emotional Spontaneous Speech Task

In this work, we present an approach to understand the computational methods and decision-making involved in identification of emotions spontaneous speech. The selected task consists Spanish TV debates, which entail a high level complexity as well additional subjectivity human perception-based annotation procedure. A simple convolutional neural model is proposed, its behaviour analysed explain decision-making. proposed slightly outperforms commonly used CNN architectures such VGG16, while being much lighter. Internal layer-by-layer transformations input spectrogram are visualised analysed. Finally, class visualisation interpretation whose usefulness assessed work.