There are a lot of questions about HTDemucs that come up regularly: which model variant is actually best, does the bitrate of your source file matter, how clean is the reconstruction when you sum the stems back. I ran a set of five structured tests on the MUSDB18-7s benchmark dataset to get actual numbers rather than subjective impressions. All tests ran locally on Apple M4.
The full data tables are on the research hub. This post summarises the key findings.
Which HTDemucs model is best
Short answer: HTDemucs (base) achieved the highest mean SDR at 8.38 dB across all four stems on the test set. It averaged 1.8 seconds per 7-second track on M4.
The speed range across models was notable: HTDemucs 6S (6-stem) processed tracks in 1.6s on average, while HTDemucs FT (fine-tuned) took 6.4s. For most producers the quality difference justifies the time cost, but batch processing scenarios change that calculation.
Full model comparison: /research/model-comparison/
Does input format actually matter
Yes, measurably. MP3 128kbps input produces stems with a mean SDR of 7.8 dB, compared to 8.04 dB for 24-bit WAV. That’s a 0.24 dB difference. The vocal stem shows the most degradation, which makes sense given how MP3 compression handles mid-frequency content.
For most production use, MP3 320kbps is close enough to lossless that the quality difference is small. Below 192kbps the degradation becomes more consistent.
Full format comparison: /research/format-quality/
How much information does separation actually lose
In 4-stem mode, summing the separated stems back produces a reconstruction that correlates with the original at r = 0.996. The difference signal sits at -21.6 dB below the original, meaning there’s real but limited information loss in the separation process.
6-stem separation shows higher reconstruction error, which is expected: dividing the signal into more components means more rounding and leakage at each split.
Full reconstruction data: /research/reconstruction-fidelity/
What predicts separation quality
The strongest correlate of vocal SDR across the test tracks was Chroma variance (harmonic complexity) (Pearson r = 0.522). In plain terms: tracks where the harmonic content is clearly separated from percussive content in the frequency domain tend to separate more cleanly. Dense arrangements with a lot of frequency overlap between instruments are harder for the model to disentangle.
Full analysis: /research/complexity-prediction/
All results
All five tests, methodology notes, and full data tables are on the research hub. Tests ran on March 14, 2026 using MUSDB18-7s (the 7-second sample of the standard MUSDB18 benchmark dataset) on Apple M4 with MPS acceleration. I’ll re-run and update these when HTDemucs releases a major model update.