Biomedical signals are frequently contaminated by colored noise; consequently, noise recognition and reduction are critical to biomedical systems. Conventional techniques have not been sufficiently focused on noise classification and using dominant noise, which facilitates noise recognition and reduction. In addition, the dependence of previous methods on threshold and calibration parameters decreases the performance of noise reduction methods. Hence, this paper enriches empirical mode decomposition (EMD) by dominant noise and presents an adaptive denoising method based on deep learning. The proposed method first identifies the dominant noise using mode decomposition and a two-step long short-term memory (LSTM) deep classifier. Then, detected intrinsic features are used for noise-aided mode decomposition. Finally, denoising is adaptively performed using the most relevant components to the dominant noise. This method is capable of accurately classifying and suppressing white and colored noise. The proposed method is validated by ECG and audio datasets. The evaluation results show that the suggested method leads to a promising improvement in noise classification, moreover, noise reduction is superior to the conventional methods in terms of SNR and RMSE criteria. As a result, the suggested method can impact noise reduction performance in real-world applications by utilizing dominant noise detection.