Cross-class zero-shot audio classification enables recognition of unseen-class audio via descriptions for connecting seen and unseen classes, which can reduce the dependence on extensively labelled data. Nevertheless, current works on zero-shot audio classification confront challenges: The large gaps between the descriptions and real-world audio samples, and the failures for textual descriptions on characterising auditory perception. To address these issues, we propose a cross-class zero-shot audio classification approach via generative learning on reconstructed auditory description. The proposed approach integrates two key modules: One is a auditory description reconstruction module, obtained via an auditory confusion matrix for aligning with acoustic features. The other is a generative learning module comprising generative adversarial networks for generating unseen-class audio. Afterwards, experimental results on audio datasets indicate that, the proposed approach outperforms state-of-the-art zero-shot learning approaches on cross-class zero-shot audio classification.