Steerable discovery of neural audio effects(csteinmetz1.github.io)

47 pointsby ushakov5 tahun yang lalu12 comments

csteinmetz1.github.io

Steerable discovery of neural audio effects

https://csteinmetz1.github.io/steerable-nafx/

13 comments

danuker5 tahun yang lalu

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.

ushakov5 tahun yang lalu

music is so overcompressed nowadays even an advanced algorithm like this will choke

nomel5 tahun yang lalu

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/

erik_p5 tahun yang lalu

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)

ushakov5 tahun yang lalu

> how can I make my mix louder?

https://www.youtube.com/watch?v=s_ANEQu5Lto

nomel5 tahun yang lalu

A trick could be to have the first 30 seconds or minute of your song at a low amplitude. This will force them to manually increase the volume of their device so they can hear anything. After that initial quiet phase of your song, you can ramp up the volume to the original level! More louder!

erik_p5 tahun yang lalu

That was brilliant, thanks.

BBC-vs-neolibs5 tahun yang lalu

bryzaguy5 tahun yang lalu

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.

ushakov5 tahun yang lalu

longer samples and longer training time

ushakov5 tahun yang lalu

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

jareklupinski5 tahun yang lalu

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?

erik_p5 tahun yang lalu

They should nickname this "clone-a-tone"