Analýza zvukových záznamů dětského pláče pomocí metod strojového učení

This bachelor thesis focuses on the analysis of audio recordings of infant crying using machine learning methods and on the design of a web application for its automatic classification. The aim of the thesis is to verify the extent to which selected physiological states of an infant can be distinguished based on the acoustic features of an audio signal and to use these findings in the implementation of a functional software prototype. The theoretical part of the thesis summarizes the medical, biological, and acoustic context of infant crying, the principles of digital audio signal processing, and current approaches to infant cry classification using machine learning. The thesis also includes an overview of available datasets and existing applications focused on the interpretation of infant crying. The experimental part of the thesis uses the publicly available Donate-a-Cry dataset, which contains audio recordings of infant crying divided into five classes. Acoustic features based primarily on Mel-frequency cepstral coefficients were extracted from the recordings, and different classification algorithms and strategies for addressing class imbalance were subsequently compared. In total, 27 experimental configurations were evaluated, including logistic regression, Support Vector Machine, and Random Forest. Based on the achieved results, a Random Forest model with a basic feature set and algorithm-level balancing was selected for implementation. This model achieved an accuracy of 86.96%, an MCC value of 0.432 and a macro F1-score of 0.469. The results indicate that automatic classification of infant crying based on acoustic features is possible; however, its reliability is significantly affected by the limited size of the dataset, class imbalance and the subjectivity of annotations. The output of the thesis is a web application that enables the classification of infant crying from an audio file or a direct microphone recording. The application represents an experimental prototype and is not intended for medical diagnosis. The contribution of the thesis lies in the transparent design and evaluation of a classification procedure based on publicly available data and in its practical implementation in a web environment.