Comparative Analysis of Classifiers using Wav2Vec2.0 Layer Embeddings for Imbalanced Stuttering Datasets

Stuttering is a speaking problem affecting normal fluency and flow of speech. Despite research done in this area, stuttering detection and identification is challenging in nature, due to the heterogeneity of audio samples which varies across people who stutter. This paper presents the application of extracting feature representations using pretrained deep learning models i.e. Wav2Vec2.0 (BASE & LARGE) and examines different data sampling methods for effective analysis. We have explored different sampling techniques i.e. oversampling and undersampling and evaluated their performance on traditional classifiers i.e. k-Nearest Neigh-bors (kNN), Support Vector Machines (SVM), Naïve Bayes' Classifier (NBC), Random Forest (RF) where SVM performed the best. With this, we performed a comprehensive layer-wise analysis of Wav2Vec2.0 feature embeddings and observed that middle layers capture more stuttering information. Our results show significant improvements in unweighted average recall (UAR) using the KSoF [1] and Sep-28k [2] datasets. Specifically, we observed a relative increase of 15.5% & 24.4% in UAR for the KSoF and Sep-28k dataset, relative to the results reported in [3] and [4] respectively.