Text this: Audio magnetotelluric denoising using time–frequency joint feature learning