Deep Learning Based Speech Enhancement

Abstract

This project focuses on exploring the effectiveness of deep learning systems in improving speech quality. The approach employs a fully attention-based mechanism that utilizes deep learning to enhance speech signals by processing noisy speech signals and producing perceptually enhanced clean speech signals. The model is trained on a large dataset of both noisy and clean speech signals and evaluated using both objective and subjective metrics on different benchmark datasets. Results show that the proposed method outperforms traditional speech enhancement techniques in terms of speech quality and intelligibility. The study also investigates the impact of various architectural and training parameters on the model's performance, demonstrating the potential of deep learning-based speech enhancement using Transformers-based forward feed models as a promising research area.