Traditional audio steganography by cover modification causes changes to the cover features during the embedding of a secret, which is easy to detect with emerging neural-network steganalysis tools. To address the problem, this paper proposes a coverless audio-steganography model to conceal a secret audio. In this method, the stego-audio is directly synthesized by our model, which is based on the WaveGAN framework. An extractor is meticulously designed to reconstruct the secret audio, and it contains resolution blocks to learn the different resolution features. The method does not perform any modification to an existing or generated cover, and as far as we know, this is the first directly generated stego-audio. The experimental results also show that it is difficult for the current steganalysis methods to detect the existence of a secret in the stego-audio generated by our method because there is no cover audio. The MOS metric indicates that the generated stego-audio has high audio quality. The steganography capacity can be measured from two perspectives, one is that it can reach 50% of the stego-audio from the simple size perspective, the other is that 22-37 bits can be hidden in a two-second stego-audio from the semantic. In addition, we prove using spectrum diagrams in different forms that the extractor can reconstruct the secret audio successfully on hearing, which guarantees complete semantic transmission. Finally, the experiment of noise impacts on the stego-audio transmission shows that the extractor can still completely reconstruct the semantics of the secret audios, which indicates that the proposed method has good robustness.
