Speech Enhancement Using Deep Neural Networks

Abstract

A comprehensive study is conducted to enhance audio quality in challenging noisy environments, departing from conventional approaches that target specific sound components. This paper focuses on a modified U-Net architecture integrat- ing broader audio features and implementing a probabilistic framework for direct spectral content reconstruction. Multiple variants of this system were rigorously tested across diverse noise levels and reverberation conditions, with performance evaluation conducted using objective metrics such as SDR, signal-to-noise ratio, evaluation of voice, and intelligibility scores. The paper demonstrates that proposed enhanced U-Net architecture, characterized by strategically designed connections within its structure, consistently outperforms traditional audio enhancement methods across a range of noise scenarios. Notably,the improvements in audio quality were most pronounced in highly reverberant environments, where conventional techniques often struggle to deliver satisfactory results. These results high- light the effectiveness of our novel approach in significantly enhancing audio fidelity and intelligibility, particularly in real- world noisy conditions.