Guide
How We Test Stem Splitters
See the source audio, exact model configurations, export settings and scoring window behind the RemixStem backing-track comparison, including its limitations.
By RemixStem ยท Updated
Start with a backing track whose contents are known
We used the vocals, drums, bass and other reference files from the TorchAudio music separation tutorial. Each is five seconds of stereo audio at 44.1 kHz. We summed the four files to make the input. We also summed drums, bass and other to make the target backing. Both received the same gain, chosen to put the input peak at 0.85 of full scale.
Run each method on the same input
- Demucs ran the four-stem HTDemucs model with zero random shifts. For the backing, we subtracted the estimated vocal from the mix.
- TorchAudio used HDEMUCS_HIGH_MUSDB_PLUS and the same vocal-subtraction rule. These two runs are related implementations with different weights.
- Open-Unmix used umxhq with vocals as its target and the remaining signal as its residual output.
- Spleeter 2-stem FP16 and UVR_MDXNET_9482 ran on CPU through sherpa-onnx. We kept the accompaniment output of each.
- The librosa run selected repeating patterns per channel using a cosine nearest-neighbour filter with a two-second exclusion width, then applied a soft mask. FFmpeg used half the left-minus-right signal on the left and its inverse on the right. These provide two simple alternatives to compare with a trained separator.
Score the common section of every export
We saved stereo floating-point WAVs without normalising individual outputs. We scored 0.25 to 4.70 seconds in every file so filter startup and slightly different export lengths would not give any method extra samples. For low-end level we applied the same fourth-order Butterworth bandpass from 40 to 150 Hz to both channels and compared the sum of squared samples to the reference.
For each reference/error score, we divided reference energy by the energy of the difference between output and reference and expressed that ratio in decibels. We also scored the untouched mix. There was no gain fitting, manual alignment adjustment, loudness rating or listening panel. A matching bass level can coexist with audible vocal bleed. The results table includes both the full-band error and the low-band measurements so readers can see the difference.