Guide
Best Stem Splitter for a Backing Track? Seven Methods Tested
Compare seven local separation methods on the same recording. See which preserved the bass, how closely each matched the backing, and what to try on your song.
By RemixStem · Updated
On our five-second stereo excerpt, TorchAudio Hybrid Demucs and Demucs HTDemucs produced the closest overall match to the known backing. The full-band scores were 18.62 and 18.55 dB respectively. A gap of 0.07 dB on one excerpt gives little reason to choose between them. For a singer starting out with a local command-line tool, Demucs has a useful shortcut that writes the vocals and backing as two files. Follow the vocal-removal walkthrough.
What came out of each run
The first five rows are neural configurations, including two generations of the Demucs family. The next two use signal processing. The final row shows how the original mix scores if you leave the singer in. All seven methods ran locally on the same audio.
| Method | Low-end level change | Low-band reference/error | Full-band reference/error |
|---|---|---|---|
| Demucs HTDemucs | −0.01 dB | 30.68 dB | 18.55 dB |
| TorchAudio Hybrid Demucs | 0.00 dB | 29.92 dB | 18.62 dB |
| Open-Unmix umxhq | −0.01 dB | 28.25 dB | 16.88 dB |
| Spleeter 2-stem FP16 | −0.02 dB | 28.56 dB | 16.41 dB |
| UVR MDX-Net 9482 | −0.01 dB | 29.58 dB | 17.51 dB |
| librosa repeating-background mask | −1.58 dB | 8.78 dB | 7.28 dB |
| FFmpeg stereo centre cancellation | −20.21 dB | 0.04 dB | 0.31 dB |
| Original mix, for reference | +0.01 dB | 27.44 dB | 10.29 dB |
Low-end level compares energy from 40 to 150 Hz against the actual instrumental. Zero means matching energy. The two reference/error columns compare the audio waveforms; higher means closer to the backing. The untouched mix matched the bass level within 0.01 dB even though it still contained every vocal. Read the full-band column before deciding whether a backing track improved.
Which one would we use?
For singing practice, start with Demucs and check the loudest chorus in your own song. Both Demucs configurations improved the full-band score by more than 8 dB over leaving this mix alone. If all you need is the backing, skip splitting out individual drum and bass files.
For keeping the bass, all five neural runs stayed within 0.02 dB of the reference low-band level. Centre cancellation lost about 20 dB in that band. If a karaoke export sounds thin, try a two-stem model before trying to restore the missing notes with EQ.
For a full song, audition a verse, a chorus and a transition. This comparison covers one short excerpt, CPU inference and stereo WAV exports. It cannot rank listening quality across genres, upload speed or commercial web services. For the sample, model settings and measurement window, see how we test.