CNN-Based U-Net Deep Learning Model for Speech Signal Enhancement

Abstract: Speech signals are often corrupted by environmental noise during acquisition and transmission, which degrades speech quality and intelligibility and adversely affects downstream applications such as speech recognition, communication systems, and assistive hearing devices. Traditional speech enhancement techniques, including spectral subtraction and Wiener filtering, rely on statistical assumptions about noise and frequently fail under non-stationary and real-world noise conditions. To overcome these limitations, this paper proposes a CNN–U-Net–based deep learning framework for speech enhancement that effectively suppresses noise while preserving essential speech characteristics. In the proposed approach, noisy speech signals are first transformed into the time–frequency domain using the Short-Time Fourier Transform (STFT). A U-Net architecture enhanced with convolutional neural network (CNN) layers is then employed to learn a nonlinear mapping between noisy and clean speech representations. The encoder–decoder structure captures both local spectral patterns and long-range contextual information through skip connections, enabling accurate reconstruction of clean speech components. The enhanced speech signal is finally obtained using the inverse STFT. The performance of the proposed framework is evaluated using real-world noise conditions, including traffic, fan, and household noises, at different signal-to-noise ratio (SNR) levels. Quantitative evaluation using metrics such as SNR improvement and Mean Squared Error (MSE) demonstrates that the CNN–U-Net model significantly outperforms conventional speech enhancement methods. The experimental results confirm the effectiveness and robustness of the proposed approach for speech enhancement in challenging noisy environments.