Separation of singing voice from background musical noise using modified NMF and Filtering

Separation of voice from musical background is useful in many applications such as speaker identification, speaker specific information retrieval, word recognition etc. where background music is considered as a noise. Although speech separation has been extensively studied from last many years, sufficient work is not done on voice or speech separation from background musical noise. We propose a system to separate speech from musical background. Our system consists of two stages, in which modified nonnegative matrix factorization is used to decompose the input mixture spectrogram. Discontinuity thresholding is applied on the mixture spectrogram to select out NMF components. This discontinuity is considered in temporal (time) and spectral (frequency) direction. These selected NMF components are then eliminated from the mixture and filtering is applied after resynthesization. Extensive testing on MIR-1K public dataset is done.