Spleeter vs HTDemucs: Quality and Speed on Apple M4

Head-to-head comparison on 50 MUSDB18-7s test tracks -- SDR quality and processing time

Run: March 15, 2026 Dataset: MUSDB18-7s Tracks: 50
Key finding See results table for SDR and speed comparison.

Background

Spleeter was released by Deezer Research in 2019 and became one of the most widely used open source stem splitters. It uses a U-Net convolutional architecture trained on TensorFlow. HTDemucs arrived later, in 2022, combining a hybrid transformer-convolutional architecture that significantly raised the state of the art on MUSDB18 benchmarks.

Both the 2022 HTDemucs paper (Defossez) and subsequent comparisons report HTDemucs outperforming Spleeter in mean SDR on full-length MUSDB18 tracks. What has not been benchmarked is their relative performance on Apple Silicon in 2026, where the hardware gap between the two matters: Spleeter runs TensorFlow, which has no MPS backend, so it runs CPU-only on M-series chips. HTDemucs uses PyTorch which fully supports Apple’s MPS GPU acceleration.

This test runs both models on the same 50 MUSDB18-7s clips under identical conditions, measuring both separation quality and real-world processing speed.

Methodology

Models tested:

  • Spleeter 2.1.0 with spleeter:4stems pretrained weights – TensorFlow 2.16.2, CPU only on Apple M4
  • htdemucs base model (demucs 4.0.1) – PyTorch 2.10.0 with MPS acceleration on Apple M4

Dataset: MUSDB18-7s – the 50 test split tracks (7-second clips from the standard 150-track benchmark). Same tracks used across all 5 prior benchmark tests on this site.

Metric: BSSEval v4 SDR via mir_eval. Median across tracks per stem. Standard 4-stem targets: vocals, drums, bass, other.

Device: Apple M4 MacBook Pro. Spleeter is TensorFlow-based and does not support MPS; it runs all computation on CPU. HTDemucs is PyTorch-based and runs on MPS. Processing time is wall-clock seconds per 7-second clip, measured with time.time() around the separation call only.

Notes on interpreting speed numbers

Times are per 7-second clip. A real 4-minute song contains roughly 34 such clips. The speed difference shown here compounds directly into real-world processing time. Spleeter’s CPU limitation on Apple Silicon is a structural constraint; it cannot be resolved by updating Spleeter alone, since TensorFlow currently has no MPS support for the operations Spleeter uses.

Quality and Speed: Spleeter vs HTDemucs

SDR values in dB (median across 50 tracks). Time is average seconds per 7-second clip. Spleeter: TensorFlow CPU. HTDemucs: PyTorch MPS.

Model Device Vocals SDR Drums SDR Bass SDR Other SDR Mean SDR Avg Time (s)
Spleeter (4-stem) CPU s
HTDemucs (base) MPS 8.91 10.58 9.43 4.35 8.32 1.48s

Stem-by-Stem Comparison

Delta = HTDemucs SDR minus Spleeter SDR. Positive delta means HTDemucs wins on that stem.

Stem Spleeter SDR HTDemucs SDR Delta Winner
Vocals 8.91 dB
Drums 10.58 dB
Bass 9.43 dB
Other 4.35 dB
Raw data spleeter_comparison.yml · All research data on GitHub
How to cite 
@misc{michaels2026spleeter,
  author  = {Aaron Michaels},
  title   = {{Spleeter vs HTDemucs: Quality and Speed on Apple M4}},
  year    = {2026},
  url     = {https://stemsplitter.github.io/research/spleeter-vs-htdemucs/},
  note    = {Benchmark on MUSDB18-7s}
}