Turn any song into a practice track: vocals out, click on, key shown
StemX now adds a beat-locked click track, a lead and backing vocal split, tempo and key on every full split, and WAV downloads, all in the browser with a free preview.
What we measured, what we can legally ship, and what the numbers mean.
StemX now adds a beat-locked click track, a lead and backing vocal split, tempo and key on every full split, and WAV downloads, all in the browser with a free preview.
How to find the BPM of a song with a free finder, check it by tapping, and fix the half or double answer; ours is right within 4% on 81.9% of 998 tracks.
How to remove vocals from a song online: upload, preview 60 seconds free, download the stems, plus our measured SI-SDR of 14.15 dB and where it falls short.
Key detection is right 64% of the time on GiantSteps. Withholding low-confidence answers raises that to 80% — on the 44% of tracks where a key is shown.
Our tempo detector is within 4% on 81.9% of 998 held-out tracks. Of the misses, 114 of 181 are half, double or triple the real tempo.
Every engine we tested was in tune to within 2 cents. The difference was transients: 67–77% of the punch kept, against 98–102% for the commercial engine we did not buy.
One library's weights are non-commercial, another's licence follows the network, and the PyPI package with the right name was someone else's.
SI-SDR (scale-invariant signal-to-distortion ratio) explained, with its formula — and why stems that lag by 219 samples score -13 dB unless you align first.
We went looking for a commercial licence for Demucs and MDX23C's weights and found none. Here is what that leaves, and why we would rather say so.
UVR-MDX-NET-Inst_HQ_3, HQ_4 and HQ_5 score within 0.4 dB of each other on vocals; UVR_MDXNET_Main beats all of them by 5 dB. Median SI-SDR, 12 MUSDB18 tracks.