About StemX
StemX is a stem splitter with MIDI: it splits any song into six stems and writes a MIDI file for each. StemX at stemx.app is not affiliated with STEM-X by FUSE VOID or with any STEMX cryptocurrency token.
What it does
A full split returns six stems (vocals, drums, bass, guitar, piano and other), a MIDI file for each plus one combined file, the song's tempo and, when the model is confident, its key. On request it transcribes timed lyrics and splits the lead vocal from the backing vocals.
A free 60-second preview of all six stems needs no account. One minute of input audio covers all six stems and their MIDI. The same engine is available as a REST API that spends the same minute balance.
Who it's for
Musicians and producers who need the parts of a song they don't have the multitrack for: people learning a part by ear, players who practise with the band minus their own instrument, producers who want a bassline or a drum pattern as MIDI, singers who want a karaoke version, DJs preparing edits, and developers who want stems from an API.
How it's built
Stem separation runs BS-RoFormer-SW (bs-roformer-infer, MIT; weights enerjazzer/BS-ROFO-SW-Fixed), with HT-Demucs 6s (Meta AI Research, demucs MIT) as the fallback. MIDI comes from YourMT3+ for pitched stems and ADTOF for drums. The lead and backing vocal split uses models by UVR (Anjok07 & aufr33), and lyrics come from OpenAI Whisper (MIT). Tempo and key detection: Beat This! (Foscarin, Schlüter & Widmer, CPJKU); S-KEY (Kong et al., Deezer).
Separation and transcription run on GPU servers in Google Cloud's us-east4 region (Northern Virginia, USA). Uploaded audio is stored there only until your plan's retention period ends, and we don't train models on it. Payments go through Paddle, our Merchant of Record, so we never see card details.
How we report numbers
Accuracy figures are our own measurements, published with the method. The BPM finder: Within 4% of the true tempo on 81.9% of 998 test tracks (95% CI 79.4–84.1%). The key finder: When it answers, it's right about 80% of the time. It answers on about 44% of songs.
When the model isn't confident about a key or a tempo, the result says so instead of guessing. Competitor prices on the comparison pages carry the date we checked them and a link to the source, and we don't round them in our favour. A post whose figures are out of date carries a note saying so.
We show no ratings, reviews or customer logos until real ones exist.
Who runs it
StemX is run by a sole trader. The seller's legal details are in the Terms of Service.
Contact
Email support@stemx.app for help with a result, a bill or an API integration.