I got a bit of feedback on this topic from SMP support which I think is very insightful. I am sure John would not mind me sharing with the forum. I for one will keep an eye on SMP and look forward to what John will be able to do for us in the future.


Quote
Thanks for bringing this to my attention--it's always useful to get feedback (good and bad) on our program.

To clarify, when we first released Song Master several years ago our stem separation was based on Spleeter, however, last year we greatly improved the quality of separation with a brand new algorithm. We trained our algorithm on high-resolution audio sources so I'm not entirely surprised that we didn't cleanly separate the bass from the lower quality YouTube audio. These things can be a bit finicky and work great on some tracks and not so good on others and greatly depend on the audio tracks they're trained on.

I wouldn't base the quality of a stem separator based on the quality of just one instrument on one track. I noticed that Studio One's separator is a four-track separator. Four stems are the sweet spot for separators because each stem has unique properties/signatures. For example, percussion instruments splash frequencies across the entire spectrum, vocals have formants that help to make them stand out, bass is usually monophonic and only has to share the frequency spectrum with the kick drum. Untangling the mass of frequencies produced by other instruments is challenging. We did notice some degradation in stem quality for bass, drums, and vocals when we added support for piano and guitar. Synths are particularly messy to handle and when present can really affect the stem quality.

Stem separation is still very much in the early stages of technology development and will continue to get better over time.

Best,
John Schnurrenberger
Aurally Sound


BIAB – 2025, Reaper (current), i7-12700F Processor, 32GB DDR4-3200MHz RAM, 1TB WD Black NVMe SSD, 2TB WDC Blue SSD, 1TB WD Blue, 2 TB SK NVMe, 6 TB External, Motu Audio Express 6x6