ISSE: An Interactive Sound Source Separation Editor | Hacker News Reader

ISSE: An Interactive Sound Source Separation Editor

isse.sourceforge.net

76 pointsashitlerferad19 comments

First off - what a great idea! How am I just finding out about this?

As an audio engineer, I know the first thing I want to use this on: a drum kit. Separate the cymbals from the skins!! Separate out the room sound from the direct sounds!! Get that pesky hi-hat out of the snare track!

Secondly - why only frequencies? Why not dynamics, stereo position, phase, and transient response? If I could switch the screen to show those things instead of frequency, I could isolate all kinds of other phenomena.

For example, isolate only the transients and turn them up/down, EQ them, add reverb to them, etc (drool). Non-engineers have no idea how potent this could be.

Or isolate only the instruments appearing at 30% off-center left?

There seems to be a deafening silence with audio stuff. Probably because there's no money in it, or rather the money is all held by the recording studios / big distributors, who have no incentive to share their findings. Also Google seems to rank music-software search results really low.

Other cool projects that I only found after intensive search: http://photosounder.com/ http://www.soundhelix.com/ http://www.cerlsoundgroup.org/Loris/ http://www.klingbeil.com/spear/

The guy who wrote this is at Apple now, probably working on Logic Pro; the latest versions have some note-editing features.

Interesting that this has been out for 3 years and this is the first I've heard of it. I'll have to try out the tool to separate out sources from various recordings we make for noise control projects (for example, separating out the wheel/rail contribution and engine/motor contribution from DMU/EMU train recordings, or fan contribution and transformer contribution in electrical substation recordings).

Seeing stuff on sourceforge makes me immediately distrustful of the project. Like it feels neglected, forgotten.... Buggy. Is this generally true for projects you've seen?

> Like it feels neglected, forgotten

I feel that way about most SourceForge projects, too. If you check out this project's history [1] the last commit was in 2013. Not that working software needs regular change, but as a rule of thumb it's nice to see recent commits.

[1] https://sourceforge.net/p/isse/code/ci/master/tree/

For me, neglected stuff seems like an opportunity to contribute, not something to distrust. SourceForge also allows folks to take over abandoned projects.

The stronger indicator for me is seeing that the software is written for someone's thesis. Typically that means that development stopped on graduation.

What is your current approach if you can answer? Some kind of spatial beamforming or you must use a given recorded measurement and decompose by frequency content as best as you can?

I guess this kind of machine learning that not uses total frequency spectrum but the time-domain evolution of short-time Fourier transforms (as in the spectrogram he is painting over in the demo) has a lot of potential for acoustic source separation. For instance training the machine to recognize the particular time-frequency signature of human voice and automatically filter it from a song.

Microphone array (beamforming) or sound intensity. But since those approaches typically require expensive hardware, the current approach is basically "we don't" unless the frequency characteristics are clearly distinct and we can use traditional equalization processes.

Machine learning has potential (and a lot of people have tried) but it's only as good at it's training.

This is pretty cool!

A couple if years ago, I saw a short documentary about some commercial software that was able to take the recording of an orchestra, separate the instruments and allow the user to edit the pitch and volume of an individual note an instrument was playing, but I can't seem to find it anymore.

Might anyone here know about this software ?

You're thinking of Melodyne, by Celemony Software. It's amazing.

https://www.youtube.com/watch?v=u573PyXo-pY

I recognized the beard and glasses in the first frame of the video.

Thanks!

That was an interesting video, thanks for linking it.

Two books are mentioned in the video, both of which I would like to read;

- "History of Harmonic Pythagoreanism" by Rudolf Haase

- "The Harmony of the World" by Johannes Kepler

The original title of the book by Haase, I found by jumping to the mark where it was mentioned in the original german language video, was Geschichte des Harmonikalen Pythagoreismus. Unfortunatelly, I have not been successfull in finding an English translation of the book. I could read the German version but would only understand a few select parts so I would much prefer to read it in English. Does anybody know of an English version of the book?

It's also similar to Roland's R-MIX software:

http://www.rolandus.com/products/r-mix/

Nice! I've been wanting a tool to do exactly this for years.

Trying it out now. One question, what does the "training brush" do? Everything else is intuitive so far.

Edit: Just tried this out on a song with acoustic guitar + vocals. Worked OK. There's lots of bleed between the tracks, but they are separated to a decent degree.

Folks interested in this might be interested in ManyEars and the associated hardware projects:

http://manyears.sourceforge.net/

Amazing demo. Can it be used to separate two voices in a file?

Ugh... sourceforge.net?

No, thanks.

sourceforge isn't like that any more:

https://sourceforge.net/blog/sourceforge-acquisition-and-fut...

Reply on news.ycombinator.com