Audio magnetotelluric (AMT) sounding is an essential tool for mineral exploration and subsurface electrical structure imaging. However, persistent cultural noise can severely degrade AMT data quality, particularly in mining areas with strong anthropogenic interference. In this study, we introduce an improved K-Singular Value Decomposition (K-SVD) dictionary learning method for time-domain AMT data denoising, incorporating three key innovations: (1) automatic identification of noise-contaminated segments using local kurtosis; (2) adaptive dictionary initialization combining principal component analysis with Gaussian perturbation, which accelerates convergence and avoids local minima; and (3) online atom screening and updating to maintain the noise specificity of the learned atoms. The proposed method learns noise morphology directly from raw AMT data without requiring reference stations or external training datasets. Synthetic experiments demonstrate that the method more effectively suppresses square-wave and charge-discharge noise compared with wavelet thresholding methods and conventional K-SVD, while better preserving signal morphology. Field applications further show that it significantly enhances the quality of observed data and apparent resistivity-phase curves compared with the widely used robust estimation method. The proposed method provides an efficient, reference-free solution for AMT data denoising in mining areas with strong cultural noise.