Ambience separation from mono recordings using non-negative matrix factorization

This article describes an approach to the problem of upmixing one-channel audio signals for multi-channel reproduction. The proposed method separates an ambient signal from an audio signal by computing the difference between the audio signal and a suitable approximation of the audio signal in the time-frequency domain. The approximation is derived by means of Non-negative Matrix Factorization minimizing the Kullback-Leibler divergence. A surround audio signal is generated by feeding the separated ambient signal into the rear channels. Results of extensive listening tests confirm the listeners' preference of the upmixed audio compared to unprocessed audio and competing upmix algorithms when played back on a 5.0 system.