Hierarchical Convolutional Spectral Feature Learning for Multi-Class Disfluency Classification in the SEP-28k Stuttering Speech Corpus

DOI

10.2991/978-94-6239-799-6_3How to use a DOI?

Keywords

Audio feature analysis; Convolutional Neural Network (CNN); Deep Neural Network; Mel-Frequency Cepstral Coefficients; Multiple-class categorization; SEP-28k dataset; Stuttering

Abstract

This paper presents a deep learning framework that is used to perform automatic multi-classification of stuttering on the Sep28k dataset. This approach extracts Mel-Frequency Cepstral Coefficients (MFCCs), which help represent the spectral information of the stuttered speech audio signal, whereas a convolutional neural network (CNN) is employed for multi-class classification. The audio data are partitioned into fixed 5-s segments and converted into 40-dimensional MFCC representations to ensure uniform inputs. The proposed features are then encoded, and supervised multi-class classifiers are trained with an 80:20 train–test split. The proposed model is composed of sequential 2D convolutional layers, which are supported by max-pooling and batch normalization. The Dropout Layer within a fully connected Layer to improve generalization and minimize overfitting. The Model undergoes 30 epochs of training using Adam and multi-class log loss. This experiment achieved a test accuracy of 87.9% with good precision, recall, and F1-scores. From the confusion matrix, it is demonstrated that most stuttering sub-types are correctly classified, with only small errors between similar-sounding classes. The result shows that CNN using MFCC features is proved effective in detecting the pattern in stuttering dysfluencies. Although the model is planned as a CNN–BiLSTM system, it currently uses only the CNN layer. Although BiLSTM in the future may enhance time-based learning and it will definitely improve ac-curacy.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Afrin Alam
AU  - Amritanjali Amritanjali
PY  - 2026
DA  - 2026/09/30
TI  - Hierarchical Convolutional Spectral Feature Learning for Multi-Class Disfluency Classification in the SEP-28k Stuttering Speech Corpus
BT  - Proceedings of the International Conference on Emerging Trends and Technologies in Applications of Computer Science (ICETTACS 2026)
PB  - Atlantis Press
SP  - 29
EP  - 42
SN  - 1951-6851
UR  - https://doi.org/10.2991/978-94-6239-799-6_3
DO  - 10.2991/978-94-6239-799-6_3
ID  - Alam2026
ER  -