Stuttering Detection and Classification using Self-Supervised Models for Natural and Simulated Speech Data
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Stuttering is a speech disorder that consists of disturbances in speech that occur involuntarily. These disturbances may be repetitions, prolongations, blocks, etc., in speech. People who stutter in their speech face social barriers and are often excluded from many aspects of everyday life. They are also excluded from using modern applications like voice assistants. Accurate automatic detection of stuttered speech is crucial for developing tools to diagnose and assist in therapy. Most of the traditional methods depend on generated acoustic features (MFCC, LPCC). This limits the potential to generalise them across a wide range of speakers. This work explores the use of Wav2Vec2, a self-supervised speech representation model, to detect stuttered speech automatically. It also explicitly compares end-to-end fine-tuning with embedding-based classifiers such as SVMs, incorporating class imbalance handling. Previous studies have focused on binary detection or overall stuttering accuracy. This work aims to classify individual stuttering types using Wav2Vec2 on benchmark datasets. The model distinguishes between fluent and stuttered segments effectively. It achieved an F1 score of 0.85 on LibriStutter and 0.75 on SEP-28k. Integration of SVM classifiers and SMOTE optimisation improved the SEP-28k F1 score to 0.84. The results show that this approach demonstrates strong performance across precision, recall, and F1-score, with particularly high recall for minority stutter classes. The results indicate the potential of self-supervised models in developing reliable and generalised stutter detection systems for real-world use.
Fields of Study & Subjects
Stuttering Research and Treatment
Speech Recognition and Synthesis
Voice and Speech Disorders
Publication details
- Journal / source
- International Conference on Computing, Communication, Control and Cyber-Physical Systems (I5CPS)
- Pages
- 1–6
- Publication date
- 26 February 2026
- Language
- EN