In light of the significant surge in audio data production, encompassing various types such as multimedia data, environmental samples, sensor networks, audiovisual content, Internet of Things (IoT) applications, and cloud data, efficient signal compression techniques have gained prominence. This paper presents an audio compression system utilizing a generative adversarial network (GAN). The encoder generates a latent vector with reduced dimensionality by processing the audio signal into the frequency domain and transforming it into a Mel-spectrogram. The generator network, which is trained to create high-quality signals, then utilizes this latent vector to minimize the loss function. Through iterative back-propagation and optimization methods, dynamic non-uniformly quantized optimal latent vectors are derived, enhancing the quantization of the compressed signal. Experimental results indicate that this algorithm outperforms traditional deep learning and classical audio compression techniques, showcasing higher reconstruction fidelity, increased compression rates, and enhanced audio quality, particularly in resource-constrained IoT environments.