Audio source separation with time-frequency velocities

Separating complex audio sources from a single measurement channel, with no training data, is highly challenging. We introduce a new approach, which relies on the time dynamics of rigid audio models, based on harmonic templates. The velocity vectors of such models are defined and computed in a time-frequency scalogram calculated with a wavelet transform. Similarly to rigid object segmentation in videos, multiple audio sources are discriminated by approximating their velocity vectors with low-dimensional models. The different audio sources are segmented by optimizing a harmonic template selection, which provides piecewise constant velocity approximations. Numerical experiments give examples of blind source separation from single channel audio signals.