Interactive sound soure separation

Anything you think is missing, or needs enhancement, goes here.

If you require help using Audacity, please post on the forum board relevant to your operating system:
Mac OS X
GNU/Linux and Unix-like

Interactive sound soure separation

Permanent link to this post Posted by unfa » Wed Mar 22, 2017 6:00 pm

ISSE is an opensource project that allows the user to separate individual instruments from a mixed recording by manually separating regions on spectrograms like one could do in GIMP. There's a nice demo video on the page below if you don't know what this program does:

The development seems to have stalled around 2013, but the code is there. I wasn't able to build it though.

Do you think this could be a new valuable tool for Audacity?
Posts: 16
Joined: Sat Mar 29, 2014 10:33 pm
Operating System: Linux Mint

Re: Interactive sound soure separation

Permanent link to this post Posted by Gale Andrews » Thu Mar 23, 2017 1:32 pm

One of the current Audacity developers is interested in source separation, but it is a major piece of work. Realistically, it would probably have to be funded somehow.

________________________________________FOR INSTANT HELP: (Click on Link below)
* * * * * Tips * * * * * Tutorials * * * * * Quick Start Guide * * * * * Audacity Manual
Gale Andrews
Quality Assurance
Posts: 25595
Joined: Fri Jul 27, 2007 12:02 am
Operating System: Windows 10

Re: Interactive sound soure separation

Permanent link to this post Posted by jh90s » Thu Mar 30, 2017 10:46 pm

Source separation is an interesting topic. The problem is, as Gale mentioned, that it's not an easy task. ISSE isn't too bad for what it is (it excels at noise reduction / hum removal) but vocal isolation takes a fair bit of time to get satisfactory results.
A man by the name of Jean-Louis Durrieu had written a Python script which automates the separation of vocals (it spits out the isolated vocal estimation and the attenuated vocal track estimation) and it does an impressive job at tracking the vocal harmonics and consonants in mixed recordings. For the curious, it's called separateLeadStereo and can be found on GitHub.

On the commercial side of things, you have ADX Trax Pro by Audionamix, and Hit'n'Mix by Neuratron Group. ADX is designed to do single elements, such as a lead vocal / guitar solo, whereas Hit'n'Mix tries its hand at full mixes. Hit'n'Mix is extremely promising, but the program suffers from a few issues; primarily the quality of when you solo an element. That, and it miscategorizes instruments considerably, but it usually gets the bass notes fine (probably because the bass is lower in the mix in most songs). That, and the interface is very pretty (it's an over-glorified spectrogram with colorful shapes which represent the notes as they'd appear on a normal spectrogram). When you think of it, there's really no easy way for Neuratron Group to fix the miscategorization of notes since the human voice hits all of the frequencies which instruments do. I'm assuming the algorithms categorize the instruments by frequency range, which honestly that's the only way I can think of it being done.

All in all, I would absolutely LOVE to see something native to Audacity which allows users to do source separation, but at the same time I understand it'll take a lot of time, dedication and money to get it done. Best of luck to the developer who is interested in this fascinating (and difficult) subject.
Posts: 23
Joined: Sun Feb 26, 2017 4:05 am
Operating System: Windows 10

Return to Adding Features to Audacity

Who is online

Users browsing this forum: No registered users and 3 guests