Source separation often requires real-time processing, although it is a burdensome feature to implement in small devices with limited resources. In this presentation, I propose to convert two most popular source separation frameworks, i.e. the dictionary-based and the deep learning-based ones, into their corresponding binarized versions. As the proposed algorithms are completely redefined with bitwise logics on binarized signals, we can achieve the desired efficiency. First, for the dictionary-based source separation, we explore the use of Winner Take All (WTA) hashing, with which we can encode the partial rank orders among a few randomly chosen spectral coefficients in an integer. By representing audio spectra into these integers, the algorithm can efficiently search for the nearest neighbors of a given frequency bin of the test spectrum in the hashed dictionaries (e.g. from two different musical instruments), whose counts are eventually used to calculate the Ideal Ratio Masks (IRM). Second, Bitwise Neural Networks (BNN) are introduced as an efficient neural network solution to the source separation problem. By involving the lossy binarization process of both the network parameters and the signals into the training procedure, we can achieve a fully bitwise source separation system that also enjoys the merit of deep learning. |