A Comparative Analysis of FCNN and CNN Architectures for Speech Denoising Across Diverse Noise Frequencies

Speech denoising remains a critical challenge in audio signal processing, especially under non-stationary noise conditions. While convolutional neural networks (CNNs) have been widely adopted for speech enhancement, the potential of fully connected neural networks (FCNNs) remains underexplored, particularly under frequency-varying noise scenarios. This study presents a systematic comparative analysis of FCNN and CNN architectures for speech denoising across multiple noise frequencies. Using the Common Voice dataset, we introduced diverse noise types at 8 kHz, 16 kHz, and 44 kHz to evaluate the denoising performance of both models. Experimental results demonstrate a frequency-dependent performance disparity: at 8 kHz, both models perform similarly, with CNN showing marginally higher Signal-to-Noise Ratio (SNR) and Root Mean Square Error (RMSE). At 16 kHz, CNN achieves significantly higher SNR albeit with increased RMSE, indicating a trade-off between noise suppression and spectral fidelity. At 44 kHz, CNN comprehensively outperforms FCNN, attaining superior SNR (4.80, +0.04) and lower RMSE (2.6826, –0.1556). These findings underscore the architectural advantages of CNNs in broad-frequency and complex noise environments, while revealing FCNN’s applicability in narrowband scenarios. This research highlights the necessity of frequency-aware model selection and provides novel insights into the comparative efficacy of FCNN and CNN in speech denoising.