Summary and Future Directions

We have presented a Bayesian framework in which it is possible to estimate the probability that a particular 1 or 2 of N sources are active at a single point in STFT time-frequency space, given a stereo mixture signal and the mixing parameters for the sources. This is significant because it allows us to demix up to two, rather than just one, source for each time-frequency point. This is an important advancement in signals where sources overlap, namely musical signals. Further, the system allows us to bias the probabilities in favor of which sources are more likely to be present at given frequencies, which can be especially beneficial when voices or instruments have known frequency component ranges. Currently, the system does not explicitly identify the probability that three or more sources are active, however. In the future, we will consider the prior probabilities of sources in nontrivial detail, and seek to integrate data across frames and frequency components. We will also consider an iterative approach for source estimation in which we update priors for the active sources at a given moment based on the data in local frames.