Multi-label Dysfluency Classification

Stuttering is a neuro-developmental disorder represented in 1% of the population. Dysfluency classification is still an open research question, with concerns of which feature representation or which classifier to use. Another issue, which has been neglected so far, is how to deal with audio samples that contain more than one type of dysfluency. Research has mostly preferred considering only single-labels problems, in part due to the lack of substantial multi-labels datasets. However, the FluencyBank and SEP-28K datasets are now available and contain multi-label data, which should pave the way for more research taking this aspect into account. In this paper, we give an overview of different ways to handle multi-label classification and compare them, while fine-tuning the ResNet50 network to perform multi-label dysfluency classification. We show that, fine-tuning the ResNet50, independently of the label representation, performs better than current state of the art results.