Artificial intelligence can generate images, write code, and create transcriptions. It is attacking many fronts at once, but its guarantee of success is still far away. One of the most complex fields is music. We have previously explored initiatives like MusicLM, SongR AI (with voice and lyrics), and TextToSample, but how good is the algorithm at separating instruments? The project free-music-demixer wants to find out with an automatic, free, and offline platform.
Separating the sources of a song is nothing new. In fact, eliminating the voice to create a karaoke version is one of the most common examples of separation, but Android and YouTube have removed a part of the heavy work. However, it remains a good exercise for the new algorithms. With that in mind we arrive at free-music-demixer, a source separation platform based on Open-Unmix and the pre-trained UMX-L model.
How to Separate Song Sources with free-music-demixer
The demixer works locally in the browser, and the only download it requires is the model, with an average size of 45 megabytes (I haven't seen any progress indicator, so you have to wait for the site to report its completion). The rest is a matter of selecting the song to process and clicking the button "Load audio and demix".
In a few minutes, free-music-demixer will present five audio files with its results: Bass, Drums, Vocals, Other, and Karaoke. Needless to say, the results depend on multiple factors, from the original quality of the song to the genre, through the tempo and the type of instruments present. My tests were limited, but I can say it did a more or less decent job isolating the drums and generating the karaoke version.
To finish, the site includes a demo (in case it's not possible to run the model for any reason), and a special mode that separates sources in batches, processing more than one song at a time. Its developer also warns that free-music-demixer places a little extra pressure on the CPU and RAM, and that the UMX-L license only allows non-commercial use. Link below!
Official site: Click here
