Sparse Sinusoidal Signal Representation for Speech and Music Signals

We present a Sparse representation called Fixed Dimension Modified Sinusoid Model (FD-MSM) for parametric analysis of audible signals including speech, music and mixtures. Compared with other analysis models, the proposed scheme is both pitch independent and appropriate for sparse signal representation commonly found as a favorable choice for speech enhancement and sound separation. Using the state-of-the-art Principle Component Analysis (PCA) it is demonstrated that FD-MSM signal representation is equivalent to a non-linear mapping into sinusoidal subspace which preserves those components with largest eigenvalues by projecting the signal components into the corresponding eigen-vectors. Conducting subjective experiments, we observed that the resulting signal is perceptually indistinguishable from the original ones.