MIDI Note Is on the Grid, but the Drum Sounds Late? A Sample-Onset Audit

First prove that the MIDI event is not the moving part

A note rectangle touching a grid line is only one piece of evidence. Open the event list or the finest available piano-roll position display and confirm the stored Note On—not merely the playhead graphic—lands at the intended position. Turn off groove, Swing, random timing, note repeat, arpeggiation, and timing automation for this diagnostic copy. Do not quantize the production clip in place; duplicate it and preserve the reportable failure.

The symptom belongs in this audit only when the same drum is repeatably late during playback or in an export. Use these routing gates before touching the sample:

Build a four-hit clip that survives every comparison unchanged

Create a new one-bar 4/4 clip at a comfortable fixed tempo. Put four equal-velocity, equal-length Note Ons for the suspect drum on beats 1, 2, 3, and 4. On a separate track, place an inspected short reference click at the same four positions. The reference must begin where its visible and audible attack begins; otherwise it only substitutes one unknown sample for another. Keep the two tracks separately renderable, and export every version from the same bar start through the same tail.

1. Freeze events: note number, channel, start, length, velocity, tempo, meter, and clip boundary do not change after B0.
2. Freeze identity: use one sample layer. If velocity selects different layers or round robins, disable that variation where documented or test each returned file as a separate sample identity.
3. Freeze gain for diagnosis: keep clip, pad, track, bus, and output gain unchanged. A louder candidate may seem earlier without moving any event.
4. Keep separate checkpoints: where the host permits, render the instrument dry and the full routed result. Do not compare a dry candidate with a processed baseline.

Write four coordinates instead of one vague “late” label

Zooming into a waveform can create false precision. A file may contain room noise before a stick impact; a kick may have a deliberately soft pre-attack; a stereo channel may begin before the other. Choose the musically intended onset by listening and looking under one fixed zoom and gain. Do not redefine it for the candidate that looks best.

If the host does not expose sample-accurate coordinates, report the finest honest unit it does expose and use a qualitative result such as unchanged, earlier with attack intact, or earlier but damaged. A threshold crossing is not automatically the musical onset, and a visual measurement should not be reported as greater precision than the display supports.

Candidate M: move only the sample playback-start marker

Official tools make this a real, independent variable. Ableton documents Sample Start and Sample Length in Drum Sampler and a user-defined playback region in Simpler. Apple documents start and end markers in Quick Sampler. These facts do not say where your marker should go; they show why the MIDI event and the audio region can have different starts.

1. Duplicate B0 to M. Keep the original audio file and all MIDI, envelope, pitch, warp, effect, and gain settings unchanged.
2. Listen to the source from its current marker to O. Label that span silence/noise, useful pre-attack, or uncertain. Silence detection may provide a candidate, not a verdict.
3. Move M toward O by the smallest useful amount, preview an isolated hit, then render the same four-hit bar. Never trim beyond the attack merely to make the waveform look aligned.
4. Reject the candidate if it adds a click, removes the initial transient, changes the intended timbre, or makes H1–H4 inconsistent. A fade can prevent a boundary click but can also soften the very onset under test; record it as another variable.

A passing M result means only that this chosen playback region better aligns this sample under the frozen conditions. It does not prove that the original recording was defective. A breath, flam grace note, room lead-in, or reversed swell may intentionally begin before its main peak; in those cases the musical solution may be to keep the pre-attack and document an intentional placement offset rather than delete it.

Candidate E: restore B0, then change only amplitude attack

Sample start and amplitude attack are not synonyms. Ableton defines Attack as the time for the amplitude envelope to reach its peak after the note is played. Apple likewise defines an amplitude envelope per note and separates an optional Delay before Attack from the Attack stage itself. Akai’s MPC Sample guide describes Attack as how quickly the sound fades in from the sample start point. A marker can therefore be correct while a positive envelope attack makes the beginning quieter and perceptually later.

1. Return to B0, duplicate it as E, and restore the original sample-start marker.
2. Record the current envelope type, Delay if present, Attack, velocity-to-envelope modulation, and any sample-layer-specific envelope offset. Do not assume every product exposes all of them.
3. Set only Attack to the product’s documented minimum or neutral test value. If a separate envelope Delay is nonzero, test that as its own E2 after E; do not zero both at once.
4. Render H1–H4 again. Keep E only if the intended attack becomes earlier or clearer without a new click, unnatural snap, or changed tail role. Then return Attack to B0 and confirm the symptom returns.

An envelope can change when the sound becomes perceptible without moving the MIDI or sample marker. That is why a simple waveform comparison needs both the dry source and the rendered output. If R stays fixed but the hit merely seems more immediate because E is brighter or louder, label the result perceptual change without measured onset movement; do not rewrite it as a timing correction.

Candidate P: compare dry instrument output with the documented route

Only enter this branch after the stored events, marker, and envelope have been isolated. Ableton documents that device delay compensation normally aligns delays reported by built-in devices and plug-ins, while input-monitored tracks can follow a different low-latency path. The important conclusion is not “effects always make drums late.” It is that playback, monitoring, manual track delay, external hardware, and compensation state are separate route variables.

1. Restore B0. Render the sampler’s dry output and the normal full route over identical boundaries. If both have the same R − G, stay upstream; downstream bypassing cannot explain the offset.
2. If dry aligns and the routed result does not, copy to P. Inspect manual track delay, external inserts, routing through returns or groups, latency-compensation state, and each device’s reported latency where the host exposes it.
3. Change or bypass exactly one documented stage, render again, and restore it. Bypassing an entire chain may locate a region, but it does not identify the device inside it.
4. Compare playback and offline export separately. A live monitored path may prioritize responsiveness while the compensated playback/render path remains aligned. Report both instead of forcing one setting to stand for both jobs.

Do not “repair” a route by moving every MIDI note earlier. That couples the clip to one present signal path and can make an offline render or a future compensated route early. If an intentional negative placement is ultimately the musical choice, save it as a named arrangement decision only after the technical offset is measured and the unshifted baseline remains available.

Let the first reversible result choose the fix

Sample-onset questions

Should I trim every drum sample exactly to its first nonzero sample?

No. Nonzero data can be noise, room sound, a deliberate pre-attack, or one stereo channel beginning first. Define the musical onset under fixed gain and zoom, audition the transition, and preserve useful lead-in. There is no universal waveform threshold in this audit.

Can I just drag the MIDI note earlier until it feels right?

That can be an arrangement choice, but it is a poor first diagnostic. It cannot tell whether the offset came from the file, marker, envelope, or route, and it may become early when the sound or processing changes. Keep G fixed until one layer has been identified.

Why use four hits instead of one?

One hit cannot reveal layer alternation, round robin, velocity drift, or an inconsistent render. Four equal events are still small enough to inspect individually. If all four differ despite frozen MIDI, identify the changing sample or modulation state before averaging them.

Does a longer Attack literally move the sample?

Not necessarily. In the cited sampler documentation, Attack shapes how amplitude rises after triggering. The sample can start at the same point while its beginning is attenuated, making the useful attack later or less distinct. Compare the dry source, envelope state, and render rather than treating “late” as one mechanism.

What can I test inside FingerDrum?

In the current codebase, imported or recorded pad sounds have non-destructive start/end trim plus a “Remove Leading Silence” candidate, and sampler instruments expose an Attack control. Treat the automatic suggestion as a candidate: preview it, replay the same pattern, and keep the original source available. FingerDrum does not infer which pre-attack is musically disposable. This product scope was checked against origin/main commit 19b88dec39e797c342ec8a749e1a7c811bad00a7.