Chris Aikens
It is common among music to record multiple tracks and combine them into one waveform track. Another common function that is yet to be refined is to extract one of these tracks from the complete waveform track and be able to edit it separately. In this paper written by Paris Smaragdis from Adobe Systems, an algorithm is discussed that is able to extract a single overlaid track from a waveform track. Through the use of an "assisted sound" such as a voice, whistle, or instrument as a model for what is to be extracted, the algorithm is able to distinguish what part of the sound to extract. The user studies conducted to test their algorithm have been shown to be somewhat promising. By using the exact sound originally mixed in (the perfect scenario), the algorithm successfully extracts the track. Tests were performed with different sounds, genders, and instruments to try and perform this task and it was found that pitch along with pattern is essential to extracting the data.
As a musician myself, I find this algorithm intriguing. It would be handy to be able to extract specific parts of a song and to change/edit them at my whim. To improve upon this algorithm I would suggest adding a set of prerecorded sounds/pitches/lengths to a library. The user tries to mimic the sound they hear in the music and it first gets converted into the sound pattern formed from the library. Then the program would attempt to map this sound pattern with different instruments to the original track. Perhaps adding in the allowance for a small deviation from the converted sound pattern would be helpful as well.
No comments:
Post a Comment