| Abstract [eng] |
This dissertation examines the challenge of robust audio classification in diverse acoustic environments where background noise and signal variability reduce recognition accuracy. The study aims to develop a reliable method for detecting and classifying unusual and environmental sounds in real-world scenarios. The main tasks involve comparing time–frequency features, integrating advanced deep learning frameworks, and evaluating their performance in industrial, urban, and natural soundscapes. The novelty of this work lies in its cross-domain approach using three datasets (MIMII, ESC50, and FSC22) and in the development of hybrid architectures that combine recurrent neural networks (GRU and LSTM) with modern backbones such as EffNet and SWinT. By integrating spatial representations extracted by convolutional or transformer networks with temporal modeling, the proposed models capture sequential dependencies in noisy audio scenes more effectively. The methodology employs perceptually motivated representations (mel-spectrogram, MFCC, and chroma-STFT) and a rigorous 5-fold cross-validation protocol. Experiments conducted on high-performance computing infrastructure ensure reproducibility and robustness. Performance is evaluated using accuracy, F1-score, and AUC. Results show that the combination of LSTM and SWinT, particularly with mel-spectrogram features, achieves the best performance, reaching 98.9% on MIMII, 90.8% on ESC50, and 82.8% on FSC22. |
