Constrained time-variant signal modeling for identifying colliding harmonics in sound mixtures

We present a general approach to identifying time-variant colliding harmonics in pitched steady-state monophonic sound mixtures. Each sound is described by a linear-in-parameter quasi-harmonic model which captures properly instantaneous time variations of the non-stochastic sound energy. The model parameters corresponding to colliding harmonics are estimated on the basis of multiple-solution cost function L² minimization with regularization. The major advantage against the state-of-the art methods is that no additional information about the underlying sounds in the mixture is needed. A comparative study shows that the proposed method performs significantly better than an existing algorithm for separation of monaural pitched sounds from a mixture.