Source separation of harmonic sources using active source estimation method

Abstract

This project aims to achieve source separation of multiple simultaneous instruments from an anechoic, stereo mixture. Active source estimation method incorporates existing statistical techniques for Source separation as well as perceptually significant signal features used in computational auditory scene analysis. The proposed method makes use of Spatial cues, Pitch and Amplitude modulation for separation of harmonic instruments. Short time Fourier Transform is implemented on the mixture to obtain a time-frequency representation. Degenerate unmixing estimation technique (DUET) algorithm is used to estimate the Amplitude and Phase parameters. This is then followed by Delay and Scale Subtraction Scoring (DASSS) during the first stage of separation. An autocorrelation technique is then used to identify the fundamental frequency and Harmonic to Noise ratio (HNR) as a part of stage two. Stage three involves Amplitude modulation analysis and Harmonic amplitude envelope estimation. A final reconstruction is done using the estimates from stage three to obtain individual sources. However there are limitations to this method. Improvement in Source separation can be achieved with prior knowledge of the musical score.