Sparse Sinusoidal Representation for Monaural Sound Separation

We present a Fixed Dimension Modified Sinusoid Model (FD-MSM) for parametric analysis of all audible signals consisting of speech, music and their mixtures. Compared with previous analysis models, the proposed scheme is pitch independent as well as appropriate for commonly used clustering algorithms including Vector Quantization (VQ) or Gaussian Mixture Model (GMM) due to its fixed feature dimension which in turn results in a significant reduction in search space. Evaluating proposed FD-MSM by conducting subjective experiments, we observed that the resulting signal is perceptually indistinguishable from the original.

کلیدواژه ها:

STFT ، MOS ، Mel-scale ، eigendecomposition