Speech enhancement using an improved Wave-U-Net architecture based on attention

Speech enhancement, which aims to improve the quality of speech signals and the intelligibility of speech in noisy environments, is a research area that has advanced rapidly with deep learning. However, studies in the literature have predominantly been developed for English datasets, while research focusing on Turkish data has remained quite limited. In this thesis, an Improved Wave-U-Net architecture enhanced with channel attention and self-attention mechanisms is proposed, and the model is trained with a multi-component loss function (EnhancedLoss) that jointly supervises the time and frequency domains. The study was carried out in three stages: first, the contribution of the model components was examined through an ablation analysis on Turkish data; then, the model was evaluated on the English dataset (VCTK), the Turkish dataset (TSCD) , and a mixed dataset constructed by combining these two datasets, using seven objective metrics (PESQ, STOI, ESTOI, SI-SNR, CSIG, CBAK, COVL), with the training/validation splits randomly regenerated three times and the test set kept fixed. The ablation results showed that the attention modules contributed only when trained together with the perceptual loss function. The model produced measurable improvements across all three conditions; the PESQ value increased from 1.48 to 3.54 for Turkish, from 1.29 to 2.69 for English, and from 1.36 to 3.04 for the mixed condition. In the mixed condition, the quality metrics remained close to those of the monolingual conditions. The findings support the transferability of the model developed on Turkish data to English and multilingual data, and indicate that a single model can also be used in multilingual speech enhancement studies.