> If you are interested in how well we do compared to demucs in particular, we can use the MUSDB18 dataset since that is the domain that demucs is trained to work well on. There our net win rate against demucs is ~17%, meaning we do perform better on the MUSDB18 test set. There are actually stronger competitors on both this domain and our "in-the-wild" instrument stem separation domain that we built for SAM Audio Bench, but we either match or beat all of the ones we tested (AudioShake, LalalAI, MoisesAI, etc.)
So ~20% better than demucs, better than the ones they tested, but the acknowledge there are better models out there even today. So not sure "competes against SOTA models" is right, but "getting close to compete against SOTA models" might be more accurate.
alt Hacker News
Their answer:
> If you are interested in how well we do compared to demucs in particular, we can use the MUSDB18 dataset since that is the domain that demucs is trained to work well on. There our net win rate against demucs is ~17%, meaning we do perform better on the MUSDB18 test set. There are actually stronger competitors on both this domain and our "in-the-wild" instrument stem separation domain that we built for SAM Audio Bench, but we either match or beat all of the ones we tested (AudioShake, LalalAI, MoisesAI, etc.)
So ~20% better than demucs, better than the ones they tested, but the acknowledge there are better models out there even today. So not sure "competes against SOTA models" is right, but "getting close to compete against SOTA models" might be more accurate.