[Feature Request] MP3 support about stemgen HOT 8 CLOSED

Hey! Sorry but I don't want to encourage people to separate stems from lossy formats. We need to encode to WAV when using Demucs, so it's pretty bad to use MP3 files as input. You shouldn't do this.

from stemgen.

This script is opinionated on defaulting to the highest quality possible -- it means using lossless formats (both as inputs and outputs).

from stemgen.

You're right, I didn't follow the link correctly. Thank you for your work, it's incredible! I would love to do a Blind-test on some big PA though.

from stemgen.

My guess is that extracting the stems from a lossy file isn't as accurate.

from stemgen.

It worked well for a couple of tracks that I tried. A lot of people have their music library in mp3s so it would probably be beneficial to add this option.

from stemgen.

@axeldelafosse I can't find any information that back up the fact that using lossy formats as input is bad with DEMUCS... In fact DEMUCS is trained with lossy audio:

All files from the musdb18 dataset are encoded in the Native Instruments stems format (opens new window) (.mp4). It is a multitrack format composed of 5 stereo streams, each one encoded in AAC @256kbps.

from stemgen.

Sorry but I don't have the time to argue about this. Lossy formats are bad.

The quote you shared is for the compressed dataset, surely the models are trained using the uncompressed dataset.

Splitting a song from MP3 and then encoding it in AAC is clearly impacting the audio quality. Don't do that.

from stemgen.

Thank you, really appreciate it.

Yeah to be honest I'm not sure I understand exactly how Demucs is impacting the audio quality. I'll try to dig into that at some point.

from stemgen.