This study proposes a novel approach based on the Kohonen map for denoising speech signals. In this method, noisy clusters were identified using the Kohonen map based on speech frequency and energy characteristics, while the "Minimum Statistics Noise Estimation" method was used to estimate the noise level. This approach allowed for stable results even at high noise levels. As features, MFCC was used for low noise levels, while the Log-Mel spectrogram was employed for high noise levels. Experiments were conducted at various noise levels (1%, 5%, 10%, 15%, 20%, 25% white noise), and the results were evaluated using the PESQ (Perceptual Evaluation of Speech Quality) metric. The proposed approach demonstrated that combining an energy-based criterion with frequency characteristics for identifying noisy clusters significantly improves speech quality.