Enhancing Stuttering Detection and Classification Using Wav2Vec2.0

Stuttering is a common speech disorder that disrupts the normal flow of speech, impacting the lives of those affected. An automated detection system can make their lives easier. However, challenges arise due to limited data and the diverse nature of stuttering datasets. With advancements in Automatic Speech Recognition (ASR) models, similar technologies can be used to enhance Stutter Detection and Classification. This study explores the use of the pre-trained ASR model Wav2Vec2.0 [1], experimenting with different classifiers on Wav2Vec2.0 embeddings and finetuning Wav2Vec2.0 for disfluency classification using the SEP-28K [2] dataset. The novelty of our work lies in experimenting with different Wav2Vec2.0 models with SEP-28k dataset to find out the best performing model and experimenting with 2 step approach: first binary and then multiclass classification. To address data imbalance, first binary classification is imposed for Stutter Detection, followed by disfluency classification for Stutter Classification, demonstrating an average increase of 1.81% in Unweighted Average Recall (UAR) on the SEP-28k dataset. Our best result surpasses the baseline on the SEP-28K dataset by 3.17% on Total Accuracy (TA).