Математичні моделі та методи розпізнавання природного мовлення на основі нейронних мереж

Abstract

This thesis centers around the comparison of neural network-based models for speech recognition in the presence of noise. The study reviews denoising autoencoders, CNNs, RNNs, and transformer-based architectures, and evaluates their suitability for different noise types and levels. Comparison is conducted by collecting and preprocessing dataset of noisy speech recordings and experimenting to compare the performance of these models in terms of recognition accuracy, robustness to different noise types and levels, and computational efficiency.