Singing-voice separation from monaural recordings using empirical wavelet transform

Extraction or separation of vocals from music signal is required for many applications such as MIR (Music Information Retrieval), karaoke applications, interactive music player etc. Sound mixture contains both musical (non-vocal) and non-musical (vocal) part which is having significance in different types of applications. Proposed method attempts to separate out signals by effectively decomposing signal with the help of empirical wavelet transform and then finding its repeating background to distinguish it between vocal (singing voice) and non-vocal (musical) part. Repetitiveness is a special characteristic of music and at the same time it is very less explored area for performing separation of vocals/singing voice from background music. Being most suitable for non-stationary and non-linear signals, this transform expected to give better results compared to state of art methods. Experiments are conducted on song clips from standard data set MIR-1K and on personal data set consisting of song clips from various Indian languages. Performance of the algorithm is examined by various performance measurement parameters such as signal to noise ratio (SNR), source to distortion ratio (SDR), source to artifact ratio (SAR), source to interference ratio (SIR).