Related Experiment Video
Updated: Aug 5, 2026
08:27
Image Recognition and Parameter Analysis of Concrete Vibration State Based on Support Vector Machine
Published on: January 5, 2024
Sigmatism detection from child speech spectrograms using convolutional autoencoders and support vector machines
Wojciech Pieniążek1, Maria Filipek1, Oliwia Skórzewska1
1Faculty of Biomedical Engineering, Silesian University of Technology, Roosevelta 40, Zabrze, 41-800, Poland.
Computer Methods and Programs in Biomedicine
|July 30, 2026
Summary
Convolutional autoencoders (CAEs) show promise in automatically detecting speech sound disorders like sigmatism in children. These models can distinguish between different productions of Polish sibilants, aiding in early diagnosis.
Area of Science:
- Speech-language pathology
- Machine learning in acoustics
- Computational linguistics
Background:
- Sigmatism, a speech disorder affecting sibilant sounds, is common in Polish children.
- Distinguishing between retroflex and dental sibilant productions is crucial for diagnosing sigmatism.
Purpose of the Study:
- To evaluate convolutional autoencoders (CAEs) for learning acoustic representations of Polish sibilants.
- To assess the efficacy of CAEs in automatically classifying sibilant place of articulation in children's speech.
Main Methods:
- A speech corpus of 149 children producing /tʂ̑/ and /ʂ/ was used.
- Recordings were converted to spectrograms and processed by CAE models.
- Support Vector Machines (SVMs) classified latent vectors from CAEs using 10-fold cross-validation.
Main Results:
- The best CAE configuration achieved 86.08% accuracy for /ʂ/ and 79.86% for /tʂ̑/.
- Classification performance for affricate sounds surpassed previous research findings.
- Consistent results across different CAE bottleneck dimensions suggest robust feature extraction.
Conclusions:
- CAE-based acoustic representations hold potential for automatically detecting atypical sibilant articulation in children.
- This approach is particularly promising for identifying errors in affricate sounds, a less-studied area in sigmatism research.
