By SongTools.ai
"Just notch out 2-4kHz" is some of the most repeated bad advice in home recording forums. Here's the actual reason EQ can't remove a vocal from a mix, and what can.
Search any mixing forum for "how to remove vocals with EQ" and you'll find some version of the same advice: pull up a parametric EQ, find where the vocal "sits", usually somewhere around 1-4kHz, and carve out a deep notch. Some versions get more specific, suggesting a series of narrow notches chasing the vocal's harmonics. It sounds plausible. It also doesn't work, and the reason why says a lot about what a vocal actually is.
The core mistake is treating "vocal" like it lives in one frequency range you can dial in and cut. It doesn't. A human voice in a mix typically spans:
- Fundamental pitch, roughly 80-1000Hz, the actual note being sung, lower for bass voices, higher for sopranos.
- Harmonics stacking well above that, often up past 8-10kHz, which is what gives a voice its recognizable tone and character rather than sounding like a pure sine wave.
- Sibilance and breath detail, 5-12kHz+, the "s," "t," and "f" sounds, and the airy top-end texture that makes a vocal sound present and human rather than dull.
Notch out the 1-4kHz range where a vocal feels most prominent, and you've cut maybe a slice of its presence, the fundamental and lower harmonics keep coming through untouched, and so does all the sibilance and air above your notch. The vocal is diminished, not gone. Chase it further by widening the notch or adding more of them, and you run into the second, bigger problem.
That same 1-4kHz range isn't reserved for vocals. It's one of the busiest, most overlapping parts of the entire mix:
- Guitars, the body and attack of both acoustic and electric guitar sit heavily in this range.
- Synths and keys, most lead and pad sounds have significant energy through the midrange.
- Snare and cymbals, snare crack and cymbal shimmer both reach well up into vocal territory.
- Basically anything with "presence" or "clarity", mix engineers boost this exact range on other instruments specifically because it's where the ear is most sensitive.
A notch aggressive enough to meaningfully dent the vocal is also gutting the guitar, synth, or snare occupying that same slice of spectrum at the same instant. You're not surgically removing one instrument, you're taking a chunk out of the entire mix and hoping the vocal suffers more than everything else. It doesn't; everything suffers roughly equally, because EQ has no idea which sound is a voice and which is a guitar. It only knows frequency, and frequency is shared by everyone.
This is the actual concept the "just EQ it out" advice gets wrong: frequency and source are two different things. Multiple instruments can, and constantly do, occupy the exact same frequency range at the exact same moment, fully summed together into one waveform. An EQ can only ask "how loud is this frequency range right now?" It has no way to ask "which instrument is making this sound?" Once a vocal and a guitar are both present at 2kHz in the same mixed-down file, there's no EQ move that separates "the vocal's contribution to 2kHz" from "the guitar's contribution to 2kHz." To the EQ, it's just one combined number.
This is the actual, and much harder, technical problem: recovering *which source produced which part of a combined signal*, not identifying *which frequencies to cut*. That's not something equalization was ever built to do.
This is exactly the problem real AI source separation is built to solve, and it's the whole reason a tool like Vocal Remover exists as something separate from an EQ plugin. Instead of asking "what frequency is this," a model trained on isolated vocal and instrument recordings learns the actual sonic fingerprint of a human voice, its harmonic pattern, its timbre, how it evolves over time, and identifies which parts of the combined mix came from a vocal versus an instrument, regardless of what frequency they overlap at. It's solving the source-identification problem directly instead of trying to approximate it by cutting frequency bands and hoping for the best.
That's also why source separation can give you more than "vocal gone", you can pick exactly which stems you want (vocals, instrumental, drums, bass, or other instruments), paying only for what you pick: 4 credits covers vocals and instrumental together, and each additional stem beyond that adds 2 credits. None of that is possible with an EQ notch, because an EQ was never actually separating anything, it was just turning down a range of frequencies and hoping the right instrument was hiding in it.
None of this means EQ is useless around vocals; it's just solving a different problem than removal. Carving space for a vocal to sit clearly in an already-separated vocal stem, taming harshness, or reducing (not eliminating) bleed between close-mic'd sources in a multitrack session are all legitimate, common EQ moves. The failure mode is specifically trying to use EQ to erase an entire instrument from an already-combined mix; that's a source-separation problem wearing an EQ-shaped disguise, and no amount of notching fixes that mismatch.
Q: Can a really narrow, precise notch remove just the vocal and nothing else?
A: No, narrower notches just mean you're removing less of everything, including the vocal, not removing the vocal more selectively. The problem isn't the width of the notch; it's that the vocal and other instruments share the same frequencies at the same time, so any cut affects all of them together.
Q: Why does this advice keep showing up in forums if it doesn't work?
A: It's intuitive and partially true, a subtle notch really can reduce a vocal's perceived prominence a little, especially on a simple, sparse mix. That partial success on easy cases is enough to keep the advice circulating, even though it falls apart on the majority of full-band, professionally produced tracks.
Q: Is this the same thing as the "phase cancellation" vocal-remover trick?
A: No, those are two different old techniques with two different failure modes. Phase cancellation exploits stereo panning (removing anything centered in the mix) rather than frequency, and fails for different reasons, mainly modern mixing techniques that break the "identical in both channels" assumption it depends on.
Q: Does AI separation ever struggle the way EQ does?
A: It can produce artifacts on extremely dense or unusually processed mixes, since it's making an inference rather than doing exact math. But it isn't limited by shared frequency space the way EQ is. It identifies sources by their actual sonic characteristics, so it handles overlapping instruments in the same frequency range far more reliably than any EQ-based approach ever could.
All the tools discussed in this article are available on SongTools.ai, free to explore.