Streaming service Deezer not only streams music but also sells ai models to isolate stems from mixed tracks. Even better - their "research" model is free and available publically on github.
Google search for "online ai stem" will point you to ezstems.com, ezstems.com,splitter.ai and more. All are selling the exact same thing - extracting acapellas and isolated tracks from a mixed song WITH AI.
I won't be surprised if they all are using the spleeter model through their ui, or at least based their model on the spleeter model. I wouldn't care how it works if it was free but it's not - all the "free versions" are barely functional or not functional at all. Instead of payin to potential scammers i have another suggestion.
First i found https://github.com/deezer/spleeter. The "easy three steps" procedure is:
I'm not a data scientist, not a python dude either, yet on two different machines: windows PC and an OSX macbook it just did not work for me. Somewhere in the middle of package install some dependencies cant be resolved, and once stuff is done installing - nothing is installed.
After some yelling in front of the screen i resorted to plan-b
Whats docker or how to set it up - all covered online. I'll get straight to it:
First docker image for deezer google pointed me to was spleeter-web - spleeter wrapped into a webapp ui!
Sounds awesome! Great stuff! Didn't work for me.
Tensor flow isn't able to find the model files inside, and i can't see if they're even getting downloaded by cli or anything like it in the console output. Might work for someone, didn't work for me.
Wanna go SIMPLE? - Go with the official docker image. its a bit hidden and outdated at the moment, but it works.
Once your docker is up and running
export AUDIO_IN='/spleeter/in' export AUDIO_OUT='/spleeter/out' export MODEL_DIRECTORY='/spleeter/models' docker run \ -v $AUDIO_IN:/input \ -v $AUDIO_OUT:/output \ -v $MODEL_DIRECTORY:/model \ -e MODEL_PATH=/model \ researchdeezer/spleeter separate \ -i /input/test.mp3 \ -o /output \ -p spleeter:4stems-16kHz --verbose
test.mp3 -is the filename in your
4stems-16kHz - a model for extraction. Other variants are listed in official docs i've linked earlier
You're good to go!