enguetedl.ch is a small, ad-free Swiss tool for DJs: reference mastering, six-stem separation, loudness analysis and format conversion.
It runs four tools on a track you already have: match it to a reference, split it into six instrument parts, read its numbers, or change its format. Each one takes a file you upload, up to 300 MB, and none of them writes anything back to the file you started with.
Reference mastering on enguetedl.ch is matching, and matching is two steps rather than one. First the reference track is measured: its loudness, its tonal balance, the shape of its EQ. Only after that does your track move toward it. If the mix underneath has a problem the reference does not share, matching will not find it.
No reference on hand? Auto-match works the same way, against an in-house target instead of a track you picked. That target is not a preset somebody dialled in by ear: it is built from a library of club records measured with this same engine, libebur128, the reference implementation of ITU-R BS.1770, the broadcast loudness standard. The distribution of that library is published.
Nothing in that chain is a model. Every stage is either fixed signal processing or a curve fitted once, offline, and checked into the source. The match measures your track against the reference and moves it, the auto-match target is chosen by comparing your measurements against four fixed centres, and that target is rendered as shaped noise rather than pulled out of a library as somebody else's record.
So the same track with the same settings produces the same master and the same numbers on every run. Nothing is randomised between one run and the next, and the before/after you get back is measured rather than estimated.
The last stage, right before your file leaves the kitchen, is a true-peak limiter, the pass that catches any peaks trying to sneak past full scale and presses them back down. It runs at 8x oversampling with its ceiling at -0.15 dBTP, so what comes back measures under full scale on a true-peak meter and not only on a sample-peak one.
A reference is a flavor to calibrate toward rather than a recipe to copy exactly. Matching brings your loudness and tone into the reference's range; it does not clone it, and chasing an exact match past that point burns time without making the master any better. Pick a reference you like the shape of, not the closest-sounding track you can find, and stop once it is in range rather than once it is identical.
A reference you upload has to be lossless, and that gets checked rather than trusted: the file is read and its top end inspected for the hard shelf a lossy encoder leaves behind, so renaming an .mp3 to .wav does not get it past. The check looks for the abruptness of the cliff rather than for missing content, because plenty of honest club masters carry very little above 16 kHz and a dark master has to pass. Your own track is never held to it: it gets a length check and nothing else, and whatever shape it arrives in is what the chain works with.
Stem separation on enguetedl.ch splits one finished track into six instrument files: vocals, drums, bass, guitar, piano, and other (whatever does not fit the first five: synths, percussion loops, FX). The model is htdemucs_6s, the six-source variant of Demucs. It is not undoing a multitrack that was there all along; the model is guessing, instrument by instrument, where one voice ends and the next begins.
This is the one place on enguetedl.ch where a trained model runs. htdemucs_6s is Hybrid Transformer Demucs, published by Simon Rouard, Francisco Massa and Alexandre Défossez at Meta AI (ICASSP 2023, arXiv:2211.08553). It reads two views of the same audio at once, the waveform itself and its spectrogram, and a transformer in the middle lets each view attend to the other before the six files are written. It learned what an instrument sounds like from MUSDB18-HQ, a research set of 150 tracks, plus roughly 800 further songs.
Which is also why the six do not come out equal. The authors say so themselves about this variant: guitar is okay, piano is not, with a lot of bleeding and artifacts. Vocals, drums and bass are the four-source model's original job and the ones it does best.
Expect stems clean enough to loop a vocal, pull an acappella, or mute the drums and hear what is left underneath, rather than a studio-grade, zero-bleed multitrack. A little of one instrument leaking into another's file (a whisper of hi-hat under the vocals, a shadow of bass in the guitar stem) is normal, not a failed job.
Loudness analysis on enguetedl.ch reads a track once and keeps only the numbers: integrated loudness, true peak, loudness range, peak-to-loudness, tempo, musical key, and how the energy is spread across sub, bass, mids and top end. Nothing is changed and nothing is written back to the file.
Those measurements land in your Library, the list of every track you have read. It sorts on any column, files tracks into playlists, and puts any two side by side so you can see how they would sit next to each other in a set: whether one will jump 4 dB louder than the other, whether one is brighter, whether the tempos line up.
The terms are defined further down this page, and the numbers a whole corpus of club records produces are published here, so you have something to compare your own against.
Format conversion on enguetedl.ch changes the container and the codec, and nothing else. The bit depth, the sample rate, your tags and your artwork all come out the way they went in, and the finished row tells you exactly what came out. There is one knob, the format you want, because there is nothing else here worth setting.
MP3 comes out at 320 kbps. Going from a lossless file to MP3 throws data away permanently, and going back the other way cannot put it back: an MP3 converted to WAV is a bigger file carrying the same lost detail. Convert lossless to lossy when your rig needs it, keep the lossless original.
The terms the Master report and the Library print on screen, each one defined on its own.
Your upload is processed and then destroyed, and the file it produces stays available for you to collect for 60 minutes before it goes the same way. The full account of that, and of everything else we hold, is on our privacy page.
Next: take a track to the Lab, or read what 299 club records actually measure.