Spaces:
Running
title: StemSplit
emoji: 🎚️
colorFrom: indigo
colorTo: purple
sdk: static
pinned: true
StemSplit on Hugging Face
We build StemSplit — a production stem-separation API for music apps, DJs, video editors, and ML pipelines that need fast, clean vocal / drums / bass / instrumental tracks without standing up GPUs.
This Hugging Face organization is where we open-source the measurements, models, and tooling behind the service: our benchmark suite, the converted model checkpoints we run in production, and live demos you can try in the browser.
What we ship here
| Asset | Type | Status | Description |
|---|---|---|---|
| stem-separation-benchmark-2026 | Dataset | ✅ live (v1.1) | Reproducible SDR / ISR / SIR / SAR benchmark of every popular open-source separation model on MUSDB18-HQ. 850 rows, full eval methodology and source code. |
| Music Source Separation Toolkit 2026 | Collection | ✅ live | Curated list of the open-source models a developer needs in 2026 — plus our benchmark. The fastest way to evaluate the landscape. |
| htdemucs-ft-pytorch | Model | ✅ live | HT-Demucs FT (#1 open-source vocal SDR on MUSDB18-HQ) packaged with a ready-to-deploy handler.py for Hugging Face Inference Endpoints. Returns all 4 stems. |
| htdemucs-ft-drums-pytorch | Model | ✅ live | Drums specialist sub-model. ~160 MB, ~2.6× faster than the full bag, identical drum SDR. For drum extraction, beat transcription, sample-pack creation. |
| htdemucs-ft-bass-pytorch | Model | ✅ live | Bass specialist sub-model. Same speed/quality story as drums. For bassline transcription and mix rebalancing. |
| htdemucs-ft-other-pytorch | Model | ✅ live | "Other" / instrumental specialist. Pair with the vocals model for karaoke, sample-flipping, music-bed extraction. |
| htdemucs-ft-onnx | Model | ✅ live | First complete ONNX export of HT-Demucs FT on the Hub. All 4 specialists + numpy bag aggregator in one repo. Runs in onnxruntime CPU/CoreML/CUDA/DirectML, no PyTorch required. |
| htdemucs-ft-drums-onnx | Model | ✅ live | Drums specialist ONNX. ~75% smaller / 4× faster than the full bag if you only need drums. |
| htdemucs-ft-bass-onnx | Model | ✅ live | Bass specialist ONNX. For bassline transcription, mix rebalancing. |
| htdemucs-ft-other-onnx | Model | ✅ live | Other/instrumental specialist ONNX. Pair with vocals ONNX for karaoke. |
| htdemucs-ft-vocals-onnx | Model | ✅ live | #1 open-source vocal SDR (9.19 dB) as ONNX. The defensible centerpiece for iOS/Android vocal-removal apps. |
| CoreML EP profiling + INT8 quantization | Models | 🚧 in progress | Mobile-quantized variants and CoreML benchmarks — Day 3 of the ONNX project. |
| Demo Space | Space | 🚧 in progress | Upload a track, pick a model, compare separations in your browser. |
| Live leaderboard Space | Space | 📋 planned | Community-submittable, continuously evaluated leaderboard for stem separation. |
Subscribe to this org or watch our benchmark dataset to get notified when new artefacts land.
About the StemSplit API
StemSplit gives you the same SOTA open-source models you'll find in this org, wrapped in a stable REST API with credits, queueing, and a dashboard — so you can ship a separation product without operating ML infrastructure.
- 🌐 Product: stemsplit.io
- 📘 Developer docs: stemsplit.io/developers/docs
- 🔌 API reference: stemsplit.io/developers/reference
- 📚 Guides & recipes: stemsplit.io/developers/guides
Minimal example (curl):
curl -X POST https://stemsplit.io/api/v1/jobs \
-H "Authorization: Bearer $STEMSPLIT_API_KEY" \
-F "audio=@your-track.mp3" \
-F "model=htdemucs_ft"
The model parameter accepts any of the open-source models we benchmark here —
so you can verify the quality you'll get against our public SDR numbers before
shipping anything.
What we publish open-source vs what's proprietary
Open-source (here on the Hub):
- Benchmark methodology, source code, and full results
- Model conversions and quantisations for popular inference runtimes
- Demo Spaces and reference implementations
Proprietary (the API):
- Job orchestration, autoscaling, credits, and queueing
- Customer dashboard, billing, and team features
- Optimised production inference pipeline
We use the same open-source models you can benchmark above — see What StemSplit uses internally in the benchmark dataset card for the exact quality-tier → model mapping.
How we measure ourselves
Every model we publish or use in production is evaluated with the same museval-based BSS Eval v4 pipeline, on the standard MUSDB18-HQ test split, on hardware we disclose. The full pipeline is open source — anyone can clone it, re-run it, and challenge our numbers.
Get in touch
- 🐛 Found a bug in a model card or dataset? Open a discussion
- ✉️ Business inquiries: stemsplit.io/contact
- 🐦 Follow us: announcements land here first, then go to our blog and social
This org is maintained by the StemSplit team. All artefacts are released under permissive licenses unless noted; see individual repos for details.