How accurate is StemX?
What we measured, on which data, and how wide the error bars are. Each section links the post with the full method and its limits.
Stem separation
Median SDR (BSSEval v4) across 50 tracks of the MUSDB18-HQ test set (207.8 minutes of audio), run through the same code the service uses. BS-RoFormer-SW is the engine every split starts on; HT-Demucs 6s takes over if it fails or runs out of time. MUSDB has no guitar or piano stems, so those two are scored inside "other".
| Stem | BS-RoFormer-SW | 95% CI | HT-Demucs 6s | 95% CI |
|---|---|---|---|---|
| Vocals | 11.98 dB | 10.41–13.50 | 8.65 dB | 7.30–9.98 |
| Drums | 13.30 dB | 12.18–14.11 | 9.39 dB | 8.60–10.03 |
| Bass | 12.10 dB | 10.07–13.83 | 9.13 dB | 6.40–10.89 |
| Other (guitar + piano + other) | 9.39 dB | 8.75–10.28 | 5.79 dB | 5.45–6.70 |
Run 2026-10-12 on an NVIDIA L4. The CSV is published under CC BY 4.0; the MUSDB18-HQ audio is not.
Tempo (BPM)
Within 4% of the annotated tempo on 81.9% of 998 GTZAN tracks (95% CI 79.4–84.1%). Most misses are half or double the real tempo, which the finder shows next to its answer.
Musical key
When it names a key, it is right about 80% of the time on GiantSteps. It names one on about 44% of tracks and says it isn't sure on the rest.