Hello, I found this app but I don't really know how to run it. I post it here in hopes somebody knows how to use it because it seems interesting.

Audio Super Resolution Using Neural Networks

Thank you.

 

You need to learn some basics from Python language and to install Anaconda (https://www.anaconda.com/products/distribution) and check that they have published the model checkpoints and then you can follow their the instructions from the readme.

There are also other things being developped in this domain, some of them have more easy HugginhFace/Replicate/Google Colab instances :

https://github.com/zkx06111/WSRGlow
https://github.com/mindslab-ai/nuwave
https://github.com/olvrhhn/audio_super_resolution
https://github.com/brentspell/hifi-gan-bwe/
etc...

And even already commercial projects :
https://neural.love/audio

 

Hello @shake_puig, for you and me it is completely useless at the moment or do you want to program instead of making music?

 

Last edited: Feb 8, 2023

Well, to begin with, it's an academic project with a specific set of models (so far), namely upsampling low-res speech data, if I'm reading this properly.

For audio engineering (apart from speech) we'd need to train the software to be useful. But useful it might be, ML models have been successful for instance in audio separation (Spleeter, iZotope RX et al), but they took quite a while to become usable by end users. I can see it being used to upsample/reconstruct say badly encoded lossy audio with low bitrates, or creating clearer samples out of old sample libraries.

But that's yet to come.

Similar techniques are used in video processing, DLSS and FSR and such, and they're very efficient (though they leverage the GPU for this) and come up with impressive results. I don't game that much, so I haven't got the latest hot-shit GPU, not now not ever, and these things let me run games at modern resolutions, say 4K when the GPU is mainly geared towards 1080p resolutions.

But yeah, I can see it becoming useful, but it isn't yet. And unless you're familiar with the Python ecosystem, setting this thing up might be an uphill battle with no useable result so far.

 

Why programming would be against making music ?! This is such a stupid statement.

 

Hello @jarredou, I am a little more practical there, what should a normal user do with it?
Maybe you are better off in a computer forum, where it is about programming.

 

Audio Engineering + Computing is still Audio Engineering

This could be very useful someday in the field of audio restoration..

 

This is the paper: https://arxiv.org/pdf/1708.00853.pdf

TLDR: The idea is to improve lossy compression for audio by encoding the input at a lower sample rate and predict the missing information when decoding.

Examples: https://kuleshov.github.io/audio-super-res/

Could maybe be used to help restore old recordings.

Just follow jarredou's hints if you really want to dig into this

 

So discussions about, say, scripting Reaper (or writing Reaper plugins) have no place in here?

Isn't practicality having the means of achieving ends? I fail to see why this thing even when discussed in the abstract would not be useful (i.e. practical) at some point.

These days, audio engineers/generalists have to be quite well versed with computers, so I don't think this discussion should be off the table.

 

That's exactly what HE-AAC is doing (SBR), and it's working pretty well for lower bit rates. There was already mp3PRO over 20 years ago with this approach, but it never got broad acceptance.

But trying to "improve" uncompressed audio by generating or predicting something that has never been there seems kinda useless to me.

 

Please read what the OP wrote and what he wanted to know.
Stick to the topic and don't open another discussion.

 

Did I not answer the question already? You're the one derailing things here.