Estimating single-channel source separation masks: relevance vector machine classifiers vs. pitch-based masking

Audio sources frequently concentrate much of their energy into a relatively small proportion of the available time-frequency cells in a short-time Fourier transform (STFT). This sparsity makes it possible to separate sources, to some degree, simply by selecting STFT cells dominated by the desired source, setting all others to zero (or to an estimate of the obscured target value), and inverting the STFT to a waveform. The problem of source separation then becomes identifying the cells containing good target information. We treat this as a classification problem, and train a Relevance Vector Machine (a probabilistic relative of the Support Vector Machine) to perform this task. We compare the performance of this classifier both against SVMs (it has similar accuracy but is much more efficient), and against a traditional Computational Auditory Scene Analysis (CASA) technique based on a noise-robust pitch tracker, which the RVM outperforms significantly. Differences between the RVMand pitch-tracker-based mask estimation suggest benefits to be obtained by combining both.