Deep Audio Prior for Sound Source Separation

Yapeng Tian (Summer Intern), Chenliang Xu, and Dingzeyu Li (Research Scientist)
Contacts:,, and
[ArXiv Paper]


We introduce Deep Audio Prior (DAP) to address sound source separation problem without relying on any external training data.

Comparison to NMF and RPCA methods

FilenameInput Mixed AudioGroundTruth AudioNMF: Predicted AudioRPCA: Predicted AudioOur method: DAP Predicted Audio
dog_violin Sound 1
dog_violin Sound 2
basketball_guitar Sound 1
basketball_guitar Sound 2
dog_speech Sound 1
dog_speech Sound 2