Drum Levels With Live Bass: -12 dBFS for Shared Low End

Digital full scale supplies a hard ceiling: -12 dBFS sits 12 dB below it, while -6 dBFS sits only 6 dB below. That extra 6 dB makes -12 dBFS the more disciplined starting point when live bass already occupies the low end. It is a mixing choice, not a scientific clearance threshold: no supplied source tests these two drum settings against live bass or measures a resulting gain in low-end clarity.

The distinction matters because moving from -12 to -6 dBFS roughly doubles sample-peak amplitude. A synthetic kick can become so dominant that its low-frequency content masks live bass, even when neither track clips. But the size of that effect cannot be inferred from peak level alone. MultiCalculators' example of a 90 dB saw beside an 80 dB fan yields 90.41 dB combined: the quieter source adds less than 0.5 dB to the total.

Decibels are logarithmic ratios, not quantities that can simply be added. Two equal 80 dB sources produce 83.01 dB, not 160 dB, according to MultiCalculators. For a live-bass mix, treat -12 dBFS as a conservative default to audition and measure, not as an automatic promise of clarity. Relative kick-to-bass balance, true-peak readings, and monitoring conditions still determine whether the low end coexists.

02 dB Gain Math

For a 2026 lo-fi or hip-hop session running at 44.1 or 48 kHz with live bass, I use −12 dBFS as the defensible default AI-drum-bus sample peak. The gain argument is exact: moving from −6 to −12 dBFS lowers peak amplitude by 6.0206 dB, preserving that much more headroom. I treat −6 dBFS as a render-specific exception, not an interchangeable “safe” setting.

I set both targets in linear amplitude rather than subtracting logarithmic labels. Relative to 0 dBFS, a sample peak at x dBFS has amplitude 10^(x/20). The exact target ratio is therefore one half: resetting the bus from −6 to −12 dBFS halves every sample value and reduces mean-square energy to one quarter. That is not a claim that perceived loudness or spectral clarity falls by the same proportion.

Before comparing settings, I scan both channels of the finished stereo drum bus, find the maximum absolute sample, and normalize from that measurement. A channel fader marked −6 dB is only trim relative to its control point; its position does not establish that the resulting bus peaks at −6 dBFS. Metering the rendered signal avoids making an unverified fader-setting assumption.

With MusicVAE, I preserve the generated groove before normalizing: note velocity and timing pass through the selected drum instrument before becoming audio, and that rendered performance then meets the −12 dBFS reference. A post-render fader trim is linear gain. It cannot recreate velocity-dependent changes in timbre or decay, so lowering the bus must not be confused with re-performing the pattern at different velocities.

I treat the change as broadband loss reduction, not an EQ move. Uniform attenuation lowers kick energy wherever its spectrum overlaps the live bass, but it cannot remove, narrow, or retune that energy. The kick-to-drum spectral balance also remains intact because every drum-bus component receives the same change in gain; whether the remaining overlap is perceptually troublesome still depends on the actual kick and bass.

I keep sample peak in dBFS, inter-sample peak in dBTP, and gated programme loudness in LUFS separate. dBFS defines the −6-versus-−12 comparison; dBTP tests whether reconstructed inter-sample excursions can cause summation overs; LUFS supports programme-level comparison. A programme-loudness reading cannot certify the drum bus’s sample ceiling, and a safe dBFS reading alone does not certify true peak.

The actionable test is to apply the +6 dB kick trial and compare it with the live bass in every shared low-frequency band. The higher target is eligible only if each band remains at least 12 dB below the bass. Any failed band keeps the lower, more-headroom-preserving default; a complete pass makes the higher target defensible for that render.

Published Evidence

Published loudness specifications do not nominate an internal drum-bus sample-peak ceiling. The conservative live-bass default identified in the gain section remains the defensible choice; the higher setting is an exception earned only after the prescribed shared-band test passes. Standards and streaming targets answer different questions from headroom against a live instrument.

According to EBU R128, version 4, the specified quantity is integrated programme loudness, with a stated tolerance. An AI-drum bus is neither the finished programme nor necessarily long enough to produce a stable programme-integrated result. Matching its fader to the EBU target would confuse delivery measurement with internal gain staging. The article’s conservative default therefore wins: when live bass carries the low end, EBU programme loudness does not justify increasing the drum peak reference.

According to the ITU-R BS.1770 series, K-weighted loudness is calculated through absolute and relative gates. For a short AI-generated loop, those gates and the signal’s content determine which passages enter the integrated measurement. Sample peak instead records an instantaneous digital extreme, without K-weighting or gate state. A loop’s LUFS therefore cannot be inferred from its sample peak, and a peak reference cannot be reconstructed from LUFS. Sample peak remains useful for headroom; actual shared-band comparison remains necessary for bass clearance.

According to Spotify’s current official Loudness Normalization documentation, the service uses a mastered-audio reference and applies playback gain. That mechanism operates downstream as delivery normalization. It neither selects nor validates either internal drum-bus setting, so it supplies no evidence for relaxing the conservative default when live bass is present.

According to Google’s official Groove MIDI Dataset, documented by Müller and colleagues in the GrooveNet work, paired audio and MIDI expose human timing and velocity accents. I would use those controls as a performance audit before peak normalization: compare generated attacks with the human timing and velocity gestures, preserve the accent hierarchy, and only then assess peak headroom. The dataset supports expressive fidelity, not a particular drum-bus ceiling.

Concrete action: retain a pre-normalization render alongside its timing and velocity reference, make the prescribed drum-gain trial, and compare every shared low-frequency band against the live bass. Advance to the higher bus setting only if the complete matrix test passes; otherwise, retain the conservative default.

Band Matrix: -12 dBFS Wins the Shared-Low-End Row

A drum-bus sample-peak ceiling is not one global decision; it is a band-by-band collision test. I define a shared low-end condition as non-empty kick-and-bass energy in any of the non-overlapping low-frequency bands used for the test. “Non-empty” means measurable in both signals under the same analysis floor. I aggregate energy within each band rather than reading isolated FFT-bin peaks, and I hold the render, analysis settings, and band-edge convention constant so a shared boundary cannot be counted twice. The test therefore finds actual collisions instead of assuming every kick masks every bass note.

For the shared-band row, I mark -12 dBFS as the winner. In the controlled comparison, moving from -6 to -12 dBFS reduces generated-kick level by 6 dB precisely in the contested spectrum while leaving the live-bass signal unchanged. This is not a claim that -12 dBFS is a universal drum threshold; it is the matrix result for an actual shared-band collision.

The safe exception must be tested on the render after the prescribed kick increase, not inferred from the bus label. I take the smallest bass-minus-kick margin across every observed shared band; only a minimum that clears the required separation permits the higher setting. One failed band locks the verdict to the lower default. A favorable average elsewhere cannot repair a local collision.

When live bass has no energy in any kick band, no matrix row is shared. I mark -6 dBFS as the audibility winner only in that boundary case: -12 dBFS would surrender kick level without providing a collision-specific benefit. This is not the normal live-bass default; it is the no-collision limit of the rule.

With persistent collision, neither raw peak setting is a complete solution. I retain -12 dBFS as the starting point, then alter the kick’s spectrum or envelope—moving weight away from occupied bass regions or shortening its low-frequency decay—instead of compensating with master loudness. The next action is to revise one variable, render the same arrangement, and rerun every shared-band margin. That keeps the headroom decision separate from the question of whether the generated kick itself needs redesign.

Counter-Evidence

The caveat is methodological, not directional: with live bass, the conservative drum-bus default is not a proven universal perceptual optimum. According to the provided source data, neither setting has kick-drum, bass-drum, sub-bass, or full-mix frequency-response measurements, and no supplied result states a preferred kick, bass, or combined-low-end balance. No controlled study cited in this guide holds a live bass part, kick timbre, and playback level constant while changing only the AI-drum bus from −6 to −12 dBFS. The separation rule is therefore a conservative operating threshold, not evidence that the more conservative bus peak is inherently clearer.

The analyzer limit is equally important. A 20-Hz FFT bin is a frequency-resolution interval, not an auditory critical band. Hugo Zwicker’s critical-band work explains why nearby sub-bass partials can interact across adjacent bins instead of behaving as isolated energies. The prescribed shared-band screen across the low end, applied after the +6 dB kick change, is consequently diagnostic: it identifies measurable collision risk, but it is not a complete masking model. A passing result supports the selected render; it does not predict clarity on every monitoring chain.

A processing counterexample defeats the blanket claim that the lower ceiling is automatically clearer. If the live bass is high-passed above the kick’s occupied spectrum, or if the kick ducks the bass, the processed sum can remain controlled even though the raw stems overlap. In either edge case, the higher bus-peak candidate can be acceptable when the canonical separation test is genuinely met. Processing history—not raw stem labels—determines whether that exception exists.

Sample-peak equality is not low-band energy equality. Two kicks can share the same drum-bus sample peak while differing in decay length, occupied bandwidth, and integrated energy below the live bass. Phase creates a separate variance: aligned bass and kick transients reinforce one another; opposed transients can partially cancel in the summed waveform. The bus meter can therefore report the same value while the cases behave differently. The shared-band screen must examine the processed sum, not infer low-frequency severity from a stem peak alone.

External validity stops at the monitor crossover. Monitor bass extension, subwoofer crossover, room modes, lossy playback, and listener preference can move the preferred balance, making the meter-qualified winner a defensible starting render rather than a universal listening result. Measure the fully processed sum in every shared band, retain the conservative default if the separation condition is absent or uncertain, and permit the higher candidate only when every shared band remains at least 12 dB below the live bass. This keeps the rule operational without presenting a diagnostic as a law.

Worked 48 kHz Case

The higher drum-bus ceiling is not the safer choice when live bass occupies the same low-end screening band. In this controlled case, the −6 dBFS trial earns an exception only if the kick clears the required margin in every shared 20-Hz band after the drum change. Because it does not, the defensible worked decision is −12 dBFS.

I use the following as a reproducible audit case, not as evidence of a universal loudness optimum. The timing, tuning, peak settings, and analyzer configuration are stipulated inputs:

Using a fixed tuning reference, I place the two fundamentals close enough to make their spectral interaction consequential. This is why an overall peak comparison is insufficient: isolated-stem peak differences describe transients, whereas the exception concerns what survives in each shared band after the stems combine. The +3 dB and −3 dB peak relationships are bookkeeping targets, not proof of spectral clearance.

The Hann-window analysis supplies a native-resolution view, but the decision comes from the subsequent band aggregation. I inspect both source stems and the full render because phase interaction and overlapping kick energy can change the band result. The relevant question is not whether the kick has the higher absolute sample peak; it is whether the higher transient ceiling leaves enough separation from the live bass under the required band-by-band test.

Here the −6 dBFS render misses the required separation, so I choose −12 dBFS. If the 40–60 Hz collision remains audible after that change, I shorten or filter the kick and rerun the same screen rather than restoring the whole drum bus to −6 dBFS. The −12 dBFS setting is the conservative default; it does not falsely certify that every kick recording is universally collision-free.

Five Rules

I treat shared-band occupancy, not the destination ceiling, as the working reference. A louder drum bus can pass delivery and still be wrong. The status-quo myth equates compliance with clarity: a master under the destination ceiling is not thereby clean in the low end. True-peak compliance can only force a later reduction.

Rule 1 — Occupancy: Map the AI kick and live bass across the shared low-frequency bands used in the test. If any band contains energy from both, start the AI-drum bus at a −12 dBFS sample-peak reference. That is a provisional branch, not an automatic rejection of the higher setting: proceed to the exception test rather than choosing −6 dBFS by preference.

Rule 2 — Exception: Apply the +6 dB drum-bus change, then test every shared band against the live bass throughout the performance. Choose −6 dBFS only if all remain at least 12 dB below the bass; one failed band locks the decision to −12 dBFS, even if neighboring bands pass. According to MultiCalculators, a decibel is a logarithmic ratio, not a simple amount of sound: in its 90 dB-plus-80 dB example, the quieter 80 dB source adds less than 0.5 dB to the total. That logarithmic arithmetic is a reason not to let a nearly unchanged aggregate replace the per-band result.

Rule 3 — Spectral collision: If the kick and bass fundamentals occupy the same 20-Hz band, retain −12 dBFS and repair the event: remove kick energy at the bass note or shorten the kick’s decay. Whole-bus gain lowers both without separating their fundamentals. Remeasure after the repair, but do not promote the bus merely because the collision was attenuated.

Rule 4 — Audibility repair: If −12 dBFS leaves the snare or hats more than 6 dB below their level in the −6 dBFS reference, raise only each affected component by its own measured deficit. That preserves the kick-to-bass relationship. Restoring the whole drum bus would also restore the kick and recreate the competition the lower reference was chosen to avoid.

Rule 5 — Delivery veto: If the summed mix exceeds the destination’s published true-peak ceiling—for example, −1 dBTP when that is the stated target—lower the drum bus. Treat this as a final delivery trim, not a substitute analysis. Passing the peak ceiling does not replace the shared-band decision test, reverse a failed band, or authorize −6 dBFS.

Use the rules as an ordered gate. Save the occupancy map, per-band readings, note-focused repair, component makeup, and final bus trim; when the bass arrangement changes, rerun occupancy and exception rather than inheriting a clearance obtained under a different performance.

What to do next