AI-based automatic detection of repetitions and prolongations: A performance comparison between deep neural network (DNN) and convolutional neural network (CNN)

This study compares the learning characteristics and classification performance of deep neural network (DNN) and convolutional neural network (CNN) models for automatic stuttering detection using identical acoustic features, namely, Mel-frequency Cepstral Coefficients (MFCCs). To this end, the LibriStutter synthetic speech dataset containing five speech types - fluent, sound-repetition, word-repetition, phrase-repetition, and prolongation - was used. Both models were implemented in Python using the Keras library: the DNN adopted fully connected dense layers, while the CNN utilized convolutional layers designed to capture localized temporal-spectral patterns in the MFCCs. Performance was evaluated using accuracy, precision, recall, and F1-score. The CNN classifier outperformed the DNN across all evaluation metrics, showing substantial F1-score improvements in word repetition (143%) and prolongation (81%). The CNN demonstrated lower overfitting and superior generalization due to the ability of its convolutional structure to extract temporal-spectral patterns effectively. CNNs showed structural advantages in MFCC-based stuttering detection and provided a promising foundation for developing automated fluency-assessment systems. This study contributes to understanding how neural architectures affect stuttering-recognition performance and supports the advancement of AI-assisted clinical evaluation.