StemX App VS ShellNewHandler

StemX App's answer

Every model StemX runs is licensed for commercial use, and we publish how accurate each one is. The vocal remover uses UVR's MDX-Net models; our own benchmark measured 11.43 dB SI-SDR on vocals (7-second excerpts of 12 MUSDB18 tracks, a small benchmark, and we say so). The BPM finder is within 4% of the true tempo on 81.9% of 998 test tracks, and the key finder names a key only when it's confident: it answers on about 44% of songs and is right about 80% of the time when it does. If a result isn't reliable, StemX says so rather than guessing.

StemX App's answer

You can try it before you commit: upload a song and hear a free 60-second preview of the vocals and instrumental without creating an account. The BPM and key finders are free, analyse the whole song rather than a clip, and show half and double time. Because the models are licensed for commercial use, developers and businesses can build on the API without licensing questions. Uploads are deleted automatically, after 1 hour for anonymous previews and 3–30 days for full splits depending on plan.

StemX App's answer

Singers and karaoke fans who want an instrumental of a song, producers and DJs who need acapellas, instrumentals or a track's tempo and key for remixes and sets, musicians making practice and backing tracks, and developers who want to add vocal separation or key/BPM detection to their own apps through an API.

StemX App's answer

We set out to build a stem separator we could stand behind: one that only uses models whose licences allow commercial use, and that states its accuracy openly instead of making marketing claims. That's why StemX offers two stems (vocals and instrumental), not four. We couldn't find commercially licensed weights for the models that split further, so we ship what we can legally offer and publish how well it performs.

StemX App's answer

Python and FastAPI for the API, PyTorch and ONNX Runtime for inference on NVIDIA GPUs, UVR MDX-Net for separation, Beat This! (CPJKU) for tempo and S-KEY (Deezer) for key detection, Next.js for the website, PostgreSQL, and Google Cloud (Cloud Run and GPU VMs). Payments are handled by Paddle.