An Attention-Enhanced Convolutional Neural Network for Stuttering Disfluency Detection Using Channel and Spatial Attention

Stuttering is a speech disorder characterized by disfluency patterns such as sound repetitions, prolongations, and blocks, which significantly reduces the ability of an individual's social communication and emotional well-being. Early and automated detection of stuttering disfluencies is vital for clinical evaluation and assistive therapy technologies. However, an automated classification is remaining a challenge due to severe class imbalance and acoustic variability in pathological speech. This study proposes an attention based deep learning framework for multi-class stuttering disfluency classification using three benchmark datasets: SEP-28k, FluencyBank, and UCLASS (spanning of 40 hours of audio). To address the class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied along with the acoustic data augmentations. A hybrid feature representation combining 40 Mel-Frequency Cepstral Coefficients (MFCCs), and temporal parameters (Zero Crossing Rate, Spectral Flux Onset), and Wav2Vec 2.0 contextual embeddings was extracted. These features were classified using a Convolutional Neural Network (CNN) integrated with a Convolutional Block Attention Module (CBAM) that sequentially applies channel and spatial attention. The proposed model achieved a 10-fold cross-validation accuracy of 91.2% and an overall macro F1-score of 0.84 across four target classes (Fluent, Repetition, Prolongation, and Block). Experimental results demonstrate that the proposed attention-guided architecture provides effective disfluency recognition suitable for integration into early intervention and speech therapy assistive systems.