[HN Gopher] Steerable discovery of neural audio effects
       ___________________________________________________________________
        
       Steerable discovery of neural audio effects
        
       Author : ushakov
       Score  : 41 points
       Date   : 2021-12-15 14:39 UTC (8 hours ago)
        
 (HTM) web link (csteinmetz1.github.io)
 (TXT) w3m dump (csteinmetz1.github.io)
        
       | danuker wrote:
       | I can't thank you enough!
       | 
       | I can finally DEcompress audio! I will train a compressor in
       | reverse, and I will re-remaster all the crappily-remastered
       | tracks!
       | 
       | BTW, I only buy DRM-free music, so I can do stuff like this on my
       | personal copies.
        
         | ushakov wrote:
         | music is so overcompressed nowadays even an advanced algorithm
         | like this will choke
        
           | nomel wrote:
           | The loudness war is mostly over though, from normalization
           | being built into most streaming audio services. [1]
           | 
           | 1. https://hyperbits.com/ultimate-guide-to-the-loudness-war/
        
             | BBC-vs-neolibs wrote:
             | Somebody needs to tell the troops it's over.
        
             | erik_p wrote:
             | While I'd like that to be true, I'm skeptical of that
             | outcome from streaming normalization alone.
             | 
             | All I see in production forums is how can I make my mix
             | louder?
             | 
             | You can blow by Spotify's lufs recommendation, the
             | normalizing just knocks you down, thereby squashing your
             | mix further (even less dynamic range)
        
               | ushakov wrote:
               | > how can I make my mix louder?
               | 
               | https://www.youtube.com/watch?v=s_ANEQu5Lto
        
       | ushakov wrote:
       | a year ago time-based effects such as delay and reverb were
       | considered impossible to model using machine learning networks
       | 
       | today this is reality
       | 
       | machine learning methods are more CPU-effective, perform real-
       | time, effortless to train and provide more fidelity to the
       | listener
        
       | erik_p wrote:
       | They should nickname this "clone-a-tone"
        
       | bryzaguy wrote:
       | There is a huge difference in the stereo image and brightness of
       | the modeled vs the real reverb to my ear. I wonder what it would
       | take to improve the fidelity.
        
         | ushakov wrote:
         | longer samples and longer training time
        
       | jareklupinski wrote:
       | after listening to the clean version, did Electric Guitar - Large
       | Room sound like it had a crowd or audience in the background, or
       | is it just me?
        
       ___________________________________________________________________
       (page generated 2021-12-15 23:02 UTC)