The Influence of Dataset Partitioning on Dysfluency Detection Systems

This paper empirically investigates the influence of different data splits and splitting strategies on performance dysfluency detection systems. For this, we perform experiments using wav2vec 2.0 models with a classification head as well support vector machines (SVM) in conjunction features extracted from model to detect dysfluencies. We train evaluate systems non-speaker-exclusive speaker-exclusive Stuttering Events Podcasts (SEP-28k) dataset shed some light variability results w.r.t. partition method used. Furthermore, show that SEP-28k is dominated by only few speakers, making it difficult evaluate. To remedy this problem, created SEP-28k-Extended (SEP-28k-E), containing semi-automatically generated speaker gender information for corpus, suggest splits, each useful evaluating other aspects methods detection.