Fire Truck By Ivan Ulz

Working With Fire Truck By Ivan Ulz

I first ran into this track when someone dropped a link in a Discord channel asking if I could help stem it out. The file was a standard stereo MP3, nothing special about the quality. What mattered was the arrangement. Ivan Ulz tends to layer things in a way that makes separation messy, and Fire Truck is no exception. The track sits around ambient electronic with that signature slow-building tension he does. There is a fire truck siren sample running through it, some pad work, and a bass element that gets buried pretty deep. If you are trying to isolate individual parts, you will hit the usual problems: phase issues, midrange congestion, and that siren which pans in a way that confuses most source-separation models.

Getting Fire Truck By Ivan Ulz Into a Working State

Start with the highest quality source you can find. The streaming versions from Spotify or Apple Music are fine for casual listening but absolutely useless if you need clean stems. I found a WAV rip on a forum once and the difference was night and day for the siren layer. Load it into something like iZotope RX, or if you are on a budget, use demucs via command line. The default demucs model handles the bass and pads okay, but that siren is a different story. It sits in the center frequency band and fights with the main synth lead. I tried the mdx variant and it smeared the transients something terrible. Here is where I hit the real problem. The siren has this Doppler shift effect baked in during the second act. Almost every AI stem separator interprets that frequency movement as two different sources and tries to split them apart. You end up with a ghost siren layered over the real one. I spent probably three hours on this before I figured out what was happening.

The workaround I ended up using was actually pretty simple. I ran the track through demucs first to get a rough pass on the stems, then pulled out the vocal-equivalent track since the siren sits in a similar frequency range. I used that to create a noise profile in RX, then ran a spectral de-ess-like process just on the siren frequency band. It cleaned up the ghost artifacts without nuking the original content. Took about twenty minutes once I knew what to look for. If you are doing this for content creation or editing purposes, you might also consider just sidechaining the pads against the siren stem instead of trying to perfect the separation. It is faster and sounds more natural than any AI output will, honestly. The original mix already has that pumping quality built in, so you are not fighting the artistic intent. One thing beginners keep messing up: they try to EQ the separated stems back together and wonder why it sounds hollow. The issue is that phase alignment gets destroyed during separation. Run a utility track with the original mixed stem and flip the phase on each isolated track individually while listening for where the fullness comes back. It is not perfect but it gets you closer than leaving it as-is. Usually adds maybe ten to fifteen percent back to the coherence.

There are other tools like Ultimate Vocal Remover or Lalal.ai that people swear by, and they work fine for simpler arrangements. This track is complex enough that none of them handle it cleanly out of the box. You are going to do some manual work regardless of what you start with. The siren sample itself appears to be a real recording, possibly from an European fire department based on the tone. If you need that isolated for a project and the separation is still too messy, there are public domain emergency siren libraries online. Not ideal since the original has processing on it, but sometimes good enough depending on what you need it for.