4 ms·

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 record

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).

lighttower 10y ago
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?
- taspeotis 10y ago
  > 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/ https://sourceforge.net/p/isse/code/ci/master/tree/
- ashitlerferad 10y ago
  For me, neglected stuff seems like an opportunity to contribute, not something to distrust. SourceForge also allows folks to take over abandoned projects.
- rcthompson 10y ago
  The stronger indicator for me is seeing that the software is written for someone's thesis. Typically that means that development stopped on graduation.
JorgeGT 10y ago
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.
- Anechoic 10y ago
  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.

lighttower 10y ago

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?

taspeotis 10y ago
> 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/ https://sourceforge.net/p/isse/code/ci/master/tree/
ashitlerferad 10y ago
For me, neglected stuff seems like an opportunity to contribute, not something to distrust. SourceForge also allows folks to take over abandoned projects.
rcthompson 10y ago
The stronger indicator for me is seeing that the software is written for someone's thesis. Typically that means that development stopped on graduation.

JorgeGT 10y ago

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.

Anechoic 10y ago
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.