--- license: cc-by-4.0 language: - en pretty_name: "StemSplit Stem-Separation Benchmark 2026" size_categories: - 1K Built and maintained by the StemSplit team. Source code: > [`scripts/hf-benchmark`](https://github.com/yourusername/musicai/tree/main/scripts/hf-benchmark) on GitHub. --- ## Leaderboard (median SDR per stem) | model_id | bass | drums | other | vocals | |:------------|-------:|--------:|--------:|---------:| | htdemucs_ft | 10.38 | 10.11 | 6.34 | 9.19 | | mdx_extra_q | 11.42 | 11.49 | 7.67 | 9.04 | | htdemucs_6s | 9.11 | 9.54 | 0.22 | 8.66 | | htdemucs | 9.78 | 10.01 | 6.42 | 8.53 | Higher is better. SDR is computed with the `museval` reference implementation (BSS Eval v4) on 1-second windows, exactly the protocol used by SiSEC and the SDX challenges, so these numbers are directly comparable to results reported in the literature. --- ## Configurations ### `metrics_only` (the leaderboard) A Parquet table — one row per `(model, track, stem)` — for the full 50-track MUSDB18-HQ test split. **No audio is shipped here**; MUSDB18 is a research-only corpus and we are not allowed to redistribute it. Schema: | column | type | meaning | |---|---|---| | `model_id` | string | matches `configs/models.yaml` | | `track_id` | string | folder name in `musdb18hq/test/` | | `stem` | string | one of `vocals`, `drums`, `bass`, `other` | | `sdr_median` | float | median Signal-to-Distortion Ratio (dB) | | `sdr_mean` | float | mean SDR (dB) | | `isr_median` | float | median Image-to-Spatial Ratio (dB) | | `sir_median` | float | median Signal-to-Interferences Ratio (dB) | | `sar_median` | float | median Signal-to-Artifacts Ratio (dB) | | `n_frames` | int | number of evaluation windows | | `sample_rate` | int | typically 44100 | | `duration_s` | float | input track length in seconds | | `wall_time_s` | float | inference wall time | | `rtf` | float | real-time factor (`wall_time / duration`) | | `peak_rss_mb` | int | peak resident memory of the runner process | | `peak_mps_mem_mb` | int | peak Metal allocation (Apple Silicon only) | | `host_chip` | string | e.g. `Apple M4 Pro` | | `host_unified_memory_gb` | string | unified memory size on the test machine | | `commit_sha` | string | repo SHA the run was produced from | ```python from datasets import load_dataset ds = load_dataset("StemSplitio/stem-separation-benchmark-2026", "metrics_only") df = ds["results"].to_pandas() df.groupby(["model_id", "stem"])["sdr_median"].median().unstack() ``` ### `audio_samples` (planned for v1.1) A future config will ship a small set of CC-BY-licensed clips with reference and predicted stems, so you can A/B them in the dataset viewer. Sourcing genuinely commercial-friendly 4-stem multitracks in 2026 is non-trivial; see the [v1.1 sourcing notes](https://github.com/yourusername/musicai/blob/main/scripts/hf-benchmark/configs/tracks.yaml) for the plan. --- ## Models ### v1 (this release) All four models below produced complete results on the full 50-track MUSDB18-HQ test split. | id | family | notes | |---|---|---| | `htdemucs` | Hybrid Transformer Demucs | Facebook AI's default 4-stem model | | `htdemucs_ft` | Hybrid Transformer Demucs | Fine-tuned variant, best vocal/instrumental separation in the Demucs family | | `htdemucs_6s` | Hybrid Transformer Demucs | 6-stem model (adds piano + guitar). See [Known Limitations](#known-limitations) for `other`-stem evaluation caveat. | | `mdx_extra_q` | Demucs MDX | MDX challenge winner, quantised, 4-model ensemble — best `bass` and `drums` of the v1 lineup | ### Planned for v1.1 (~next 24-72 hours) | id | family | why deferred | |---|---|---| | `mdx_net_inst_hq3` | MDX-Net | Runner had a `use_coreml=True` keyword that's not in the version of `audio-separator` we pinned. Fix is one-line; will land in v1.1. | | `bs_roformer` | Band-Split Roformer | Currently SOTA but slow on Apple MPS due to operator fallbacks (~11 hr for full test set). Will add once we run on CUDA. | | `mel_band_roformer` | Mel-Band Roformer | Same reason. | | `spleeter_4stems` | Spleeter | Legacy baseline; TensorFlow install is brittle on Apple Silicon. | See [`configs/models.yaml`](configs/models.yaml) on GitHub for exact versions and command lines. --- ## Known limitations We're shipping v1 with two caveats called out explicitly so nobody is surprised by the numbers. ### 1. `htdemucs_6s` undersells on the `other` stem The 6-stem model splits piano and guitar **out** of the `other` stem. MUSDB18-HQ's reference `other` stem includes piano + guitar mixed in. So the 6-stem model's residual `other` output is, by design, nearly empty — and the SDR comparison against MUSDB's `other` looks much worse than it actually is (0.22 dB in v1 vs ~6 dB for the 4-stem siblings). The fair comparison is to **sum the 6-stem model's `piano + guitar + other` outputs and compare that aggregate to MUSDB's `other`**. We'll do exactly that in v1.1's eval pass and republish. Until then, treat `htdemucs_6s`'s `other` row as not-meaningful and look at its `vocals`, `drums`, `bass` rows for a fair head-to-head. ### 2. `mdx_net_inst_hq3` is missing from v1 A bug in our runner's `audio-separator` integration killed all 50 separation runs for this model. The leaderboard you see here only contains the four Demucs-family models. v1.1 will include this model. --- ## What StemSplit uses internally The [StemSplit](https://stemsplit.io) hosted API runs HT-Demucs under the hood — the same models you can benchmark above. Pick a quality tier and look up its row in the leaderboard: | StemSplit tier | Model row in this benchmark | When to choose it | |---|---|---| | `FAST` | `htdemucs` | Speed-priority previews and bulk processing | | `BALANCED` (default) | `htdemucs_ft` | Best vocal separation per second of compute | | `BEST` (6-stem) | `htdemucs_6s` | When you need piano + guitar separately | In other words: **this dataset is also a benchmark of StemSplit's own quality**. We didn't add a separate `stemsplit_api` row because it would just duplicate those numbers. --- ## Use the StemSplit API If you'd rather not stand up Demucs, ffmpeg, torchcodec, and a GPU yourself, the StemSplit API ships the same models with a single HTTP call. Pricing, quotas, and the full feature set are documented in our developer portal: | Resource | URL | |---|---| | 🏠 Developer landing | [stemsplit.io/developers](https://stemsplit.io/developers) | | 📘 Getting-started docs (auth, upload, polling) | [stemsplit.io/developers/docs](https://stemsplit.io/developers/docs) | | 📑 API reference (every endpoint, every field) | [stemsplit.io/developers/reference](https://stemsplit.io/developers/reference) | | 🧩 Integration guides (Zapier, n8n, Make, Pipedream, Discord, Audacity, DJ workflows, ...) | [stemsplit.io/developers/guides](https://stemsplit.io/developers/guides) | Minimal example — submit a job, poll, download stems: ```bash # 1. Submit JOB=$(curl -sS -X POST https://stemsplit.io/api/v1/jobs \ -H "Authorization: Bearer $STEMSPLIT_API_KEY" \ -F "file=@song.wav" \ -F "stems=4") JOB_ID=$(echo "$JOB" | jq -r .id) # 2. Poll until completed while [ "$(curl -sS https://stemsplit.io/api/v1/jobs/$JOB_ID \ -H "Authorization: Bearer $STEMSPLIT_API_KEY" | jq -r .status)" != "completed" ]; do sleep 3 done # 3. Download every stem curl -sS https://stemsplit.io/api/v1/jobs/$JOB_ID \ -H "Authorization: Bearer $STEMSPLIT_API_KEY" \ | jq -r '.stems | to_entries[] | "\(.key) \(.value)"' \ | while read stem url; do curl -sSL "$url" -o "$stem.wav"; done ``` → Get an API key at [stemsplit.io/developers](https://stemsplit.io/developers). --- ## Reproducing the results Everything runs locally on a Mac with Apple Silicon — no CUDA required. ```bash git clone https://github.com/yourusername/musicai cd musicai/scripts/hf-benchmark uv venv --python 3.11 && source .venv/bin/activate uv pip install -e . # 1. Get the data (~22 GB extracted from a 21 GB Zenodo zip) python -m src.download_musdb # 2. Run every enabled model on every track python -m src.run_all --continue-on-error # 3. Score with BSS Eval v4 python -m src.eval_metrics # 4. Assemble the HF dataset (and optionally push) python -m src.build_dataset python -m src.push_to_hub --create ``` Reference wall times measured on an Apple M4 Pro (24 GB unified memory), PyTorch 2.11 with the MPS backend, for the v1 lineup: | Stage | Wall time | |---|---| | Download MUSDB18-HQ from Zenodo | 32 min | | Separate (4 models × 50 tracks) | 2 h 3 min | | Eval (museval BSS Eval v4) | 2 h 10 min | | Build dataset | < 1 s | | **Total** | **~4 h 45 min** | --- ## Why we built this We run [StemSplit](https://stemsplit.io) — a hosted stem-separation service — and we needed an honest, public, reproducible way to compare ourselves to the state of the art. So we open-sourced it. Want to ship a separation product without standing up GPU infrastructure? The same models are one HTTP call away — see the [Use the StemSplit API](#use-the-stemsplit-api) section above, or jump straight to the [developer docs](https://stemsplit.io/developers/docs) and [API reference](https://stemsplit.io/developers/reference). --- ## Licensing - This dataset (the metrics, the dataset card, and the CC-BY audio samples) is released under **CC-BY-4.0** — please cite us if you use it. - The MUSDB18-HQ audio referenced by `metrics_only.track_id` is **not** redistributed here. Download it from [Zenodo](https://zenodo.org/records/3338373) under its own terms. - Each separation model retains its own license; see the table above. ## Citation ```bibtex @misc{stemsplit_benchmark_2026, title = {StemSplit Stem-Separation Benchmark 2026}, author = {StemSplit}, year = {2026}, url = {https://huggingface.co/datasets/StemSplitio/stem-separation-benchmark-2026} } ```