Open challenge · $1000 in USDC ·
Make vocal separation faster or cleaner than anything on this chart. Prize money follows the new area you add under the frontier, so a small, fast model can win as much as a big, clean one.
Runs in this tab. Nothing leaves your device unless you copy it out.
Your file stays on this device. It runs next to HTDemucs, so you see the same ratio the board uses.
S = √(CPU × GPU)
Times faster than HTDemucs, end to end, with a full clip including the STFT. Half is CPU, half is GPU, geometric means all the way down.
Fail on a device, or run slower than 0.05× there, and that device scores 0.05×. Failing one GPU costs a model 4× faster everywhere else about 42% of its score, so well-rounded models win.
Q = mean SDR, dB
Mean of three sets of 50 clips of 30 s: musdb30 and cc30, both public and rebuildable, and private30, which only the maintainers can run. An entry must score within 1 dB of HTDemucs to count.
Each device's output also has to match the native output on a public test clip, within 0.5 dB. Fast and wrong counts as a failure.
$ ∝ new area
One unit is one doubling of speed times one dB. Your points earn the area that nobody before you covers, in order of submission. Copies earn nothing, and an earlier rival's area stays theirs.
The pot divides over all new area, including what the maintainers' own training added. That share isn't paid out, and it shrinks the more you add.