Generative adversarial networks (GANs) have been extensively used in image domain showing promising results in generating and learning data distributions in the absence of clean data. However, the audio domain, specially underwater acoustics are not yet fully explored in reporting the efficiency and applicability of GANs. We propose an audio GAN framework called ambient noise-free GAN (AN-GAN) to address the underwater acoustic signal denoising problem by removing the background ambient noise. The proposed AN-GAN can learn a clean audio generation with improved signal-to-noise ratio (SNR) given only the noisy samples from the underwater audio dataset. The simulated and real-time data collected from online available source ShipsEar, is used for the analysis and validation purpose. The comparative analysis shows an average percentage improvement of proposed AN-GAN with GAN-based and conventional statistical underwater denoising methods as 6.27% for UWAR-GAN, 227% for Wavelet denoising, 247% for EMD and 65% for Wiener technique.