ELEN E4810 - Digital Signal Processing - Project

Removing vocals from commercial tracks

What's this?

On this page, you'll find information regarding Bhargav's DSP project that he completed in the Fall of 2007. Jaime Peretzman was Bhargav's ebullient partner on this project.

Project Description

The goal was the removal of vocals from commercially-mixed audio tracks. Three different removal methods were explored:

• Bandpass filtering - This method relies on the frequency range of the human voice to remain constant across most songs.

• Stereo cancellation - This method involves canceling common frequencies between both channels of a stereo mix. The common center-channel, which the lead singer's voice usually resides in, is often cancelled. The resulting sample, however, is single-channel (mono).

• Short Time Fourier Transform and Binary masking - This (more advanced) method is a topic of ongoing research (and a part of Blind Source Separation, see PDF here), and the implementation required the picking of a single audio track, since the method can only work on a track-by-track basis. The selected track was the ever-popular Let It Be, by The Beatles.

As with most serious digital signal processing work, this project was designed and implemented in MATLAB.

Report

The project report is available here (warning, 1.1 megabyte PDF). All project assets, including MATLAB code and audio samples are available as a .zip file. You can also grab just the MATLAB code here (though it may not run without its companion files).

Acknowledgment

Bhargav and Jaime would like to thank Prof. Dan Ellis for approving such an overly ambitious project, giving us a running start and guiding us along the way.