The "delossyfiers" advent

Discussion in 'Ai for Music' started by forart.it, Apr 24, 2025.

  1. saccamano

    saccamano Audiosexual

    Joined:
    Mar 26, 2023
    Messages:
    2,060
    Likes Received:
    880
    Location:
    CBGB omfug
    The best "delossifier" I know of is to start with a proper wave file 96/88/44khz @ 32bF/24b/16b and use a LOSSLESS codec like FLAC. With all this jabber about optimizers and tons of strings of useless code, the end result is STILL A LOSSY RESULTANT FILE WITH THE SAME ARTIFACTING GARBAGE AS ANY MP3 (or other lossy codec). Sure you can minimize the artifacts, but it's still a crappy end result IMO... Until a new phenom codec is invented that compresses like MP3 (or better) but retains all the audio information (lossless) with no artifacts I will continue using FLAC. The extra space IMO, is a better compromise than having to listen to shitty MP3 encodes..
     
  2. canbi

    canbi Kapellmeister

    Joined:
    Jun 12, 2023
    Messages:
    230
    Likes Received:
    65
    yep, to the trash it goes SHIT.png
     
  3. thatruth

    thatruth Newbie

    Joined:
    Dec 21, 2015
    Messages:
    14
    Likes Received:
    0
    Yea I tried it and saw the same thing
     
  4. canbi

    canbi Kapellmeister

    Joined:
    Jun 12, 2023
    Messages:
    230
    Likes Received:
    65
    last minute - new colab dropped

    :^)

    ^made by same user that made FFmpeg
     
  5. forart.it

    forart.it Producer

    Joined:
    May 5, 2023
    Messages:
    156
    Likes Received:
    77
    Merry Christmass to all.

    A plausible delossification can be somewhat performed (read approximated) only by exploiting neural networks...

    ...that's why "someone" are still waiting for the next - Intel - OpenVINO™ AI Plugins for Audacity !
    (note: some interesting models listed in this discussion)
     
  6. forart.it

    forart.it Producer

    Joined:
    May 5, 2023
    Messages:
    156
    Likes Received:
    77
    • Funny Funny x 1
    • Interesting Interesting x 1
    • List
  7. canbi

    canbi Kapellmeister

    Joined:
    Jun 12, 2023
    Messages:
    230
    Likes Received:
    65
  8. forart.it

    forart.it Producer

    Joined:
    May 5, 2023
    Messages:
    156
    Likes Received:
    77
    Errr... what you mean ?
    upload_2026-1-4_16-21-39.png

    DR is exactly the same (8dB)...

    ...about freq cut: check the original mp3 file yourself !
     
    Last edited: Jan 4, 2026 at 3:32 PM

    Attached Files:

  9. canbi

    canbi Kapellmeister

    Joined:
    Jun 12, 2023
    Messages:
    230
    Likes Received:
    65
    Compare the spectrograms at 17.5k and above, these are not identical. High frequency mangling is the base idea of upsampling, but if it's not preserving the original harmonics pattern and dynamics, the tool level is questionable
     
  10. forart.it

    forart.it Producer

    Joined:
    May 5, 2023
    Messages:
    156
    Likes Received:
    77
    They of course CAN'T be: (high) frequencies are GUESSED by the model in the "restored" file.

    I don't know how Machinima encoded the original file, but frequency cutoff it's an mp3 "feature" (that's why we're looking for delossifiers)

    ...anyway here's a more detailed Deltawave Null Comparator spectras:

    BEFORE:
    upload_2026-1-4_20-39-13.png

    AFTER:
    upload_2026-1-4_20-39-32.png

    DELTA:
    upload_2026-1-4_20-39-54.png

    Other interesting measurements in the HTML export.
     
Loading...
Loading...