Results 1 to 4 of 4

Thread: Neural Networking

  1. #1
    Join Date
    Apr 2008
    Posts
    1,696
    Thanks
    185
    Thanked 1,068 Times in 380 Posts

    Neural Networking

    So there's a Russian chap on Github who's working on using neural network machine learning for audio processing. He's starting with noise removal and he's still got a long ways to go. But knowing what self-taught machines have accomplished recently in other areas like playing championship chess and Go, identifying habitable planets and looking for SETI signals, I remain confident that someday - maybe even in my own lifetime, a computer will be able to take a B- audience recording of Jimi and make it sound like an EX+ sound board. I'm going to contact this Russian programmer and ask him to keep that goal in mind!

  2. The Following 5 Users Say Thank You to Roland Stone For This Useful Post:


  3. #2
    Join Date
    Feb 2015
    Posts
    43
    Thanks
    11
    Thanked 132 Times in 22 Posts

    Re: Neural Networking

    Don't get your hopes up too high; audio restoration is incredibly complex and neural networks are just very powerful statistics tools. They are no silver bullet for everything, but always trained on one very specific problem. Audio restoration is not one but many different problems: wow & flutter removal, EQ matching, reverb removal, demixing, rolloff/harmonic excitation, de-distortion, dropouts reconstruction and so on... Now there already are tools for all of these issues, but few work without a lot of user input and many are in their infancy. I'm sure neural networks could improve each of these disciplines individually. I just don't expect an all-purpose audio restoration network for quite a while, at least two more decades. For some cool neural network audio action check out ISSE: http://isse.sourceforge.net/

  4. The Following 3 Users Say Thank You to hendrixfan1995 For This Useful Post:


  5. #3
    Join Date
    Apr 2008
    Posts
    1,696
    Thanks
    185
    Thanked 1,068 Times in 380 Posts

    Re: Neural Networking

    Interesting demo. I can already do a lot of that with Melodyne. I've even taken a stab at removing the vocal from Eire Apparent's "Morning Glory"! But its so labor intensive I could only do a few seconds and with marginal results that left "holes" in the audio spectrum.

    I won't hold my breath, but I still think that if a machine can self teach itself to beat every human (and most other machines) at chess, then eventually it will be able to figure out what a fair audience recording ought to sound like ideally.

  6. The Following User Says Thank You to Roland Stone For This Useful Post:


  7. #4
    Join Date
    Feb 2015
    Posts
    43
    Thanks
    11
    Thanked 132 Times in 22 Posts

    Re: Neural Networking

    Chess is just statistics really, the more moves you look ahead, the better you are. With every round the possible outcomes multiply, so you have to keep exponentially more moves in mind the further you want to look ahead. It is really a problem of storing and calculating the best "path" between as many future moves as possible. Very mathematical, easy for computers, difficult for humans.

    Taking an audience recording into soundboard quality involves so much more guessing and de-novo synthesis of missing data. Difficult for computers and humans alike.

    You can probably get similar results with melodyne, but it takes ages. With ISSE you can do separation almost instantly, without holes - the demo doesn't show its full potential.

Bookmarks

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •  
Powered by vBitty (VBTT) 4 for XBT v1.1 CUSTOM by Toolmanwill