Real-time blind audio source separation: performance assessment on an advanced digital signal processor

Many on-line blind audio source separation (BASS) algorithms have been presented so far to the scientific community, but only a few of them have been evaluated in terms of their real-time performance. In this paper we consider a well-established BASS method (oriented to voices separation) evaluating its performance in terms of separation quality allowed by a real-time embedded computing implementation, also considering novel and state-of-the-art improvements to the it. To this aim, the algorithm has been implemented and ported for real-time execution onto an advanced low-power digital signal processor targeted for complex-domain applications. The optimized embedded implementation is able to perform up to five iterations of the gradient for any input frame of data, achieving good separation levels (up to 11.8 dB of signal to interference ratio on custom recording in real environments). The proposed solution doubles the performance of a C-optimized version running on a traditional PC processor, achieving a better separation result with lower power requirements.