Audio compression is still an up-to-date topic because the demand for big data streams is rapidly increasing. Deep learning has brought up new algorithms that decrease bitrates with good perception quality. The novel approach in generative artificial intelligence is to produce new data from prior stored in network parameters, called a deep prior. The deep audio prior framework shows its success in various tasks such as inpainting, declipping, and bandwidth extension, but it has not been tested for compression. In this paper, we test this method with a prebuilt network for inpainting. Our idea of compression is based on reducing the number of time-frequency coefficients in the spectrogram while allowing the reconstruction of the original signal with high quality.