What Latency Actually Is, and How Much Is Too Much

Reviewed September 2026.

You are tracking a vocal, or a guitar part you have played a hundred times, and it will not sit. The headphones are handing your performance back a fraction behind your body, and after twenty minutes you are either singing behind yourself or fighting the delay hard enough that the take goes stiff. Somewhere in there you open a browser tab and start reading about interfaces with lower latency figures.

Leave that tab alone for an hour. Latency is a budget spent across six places, and the interface is responsible for a smaller share of it than almost anyone assumes. Most of what you are feeling was set by a menu in your own software, and a good part of the rest by a plug-in you put on the mix bus last week and forgot about.

Where the Milliseconds Actually Come From

Follow one moment of sound from microphone to headphones. The preamp is analog and takes nothing. The A to D converter takes half a millisecond to a millisecond and a half, most of it in the anti-alias filter rather than the conversion, and no setting changes it. The driver collects samples into a block before waking your software, because handing over one sample at a time is not how any of this works. That block is the buffer setting. The interface's own transfers across the bus are a separate and smaller thing, fixed by the connection rather than by you. Then add time on the wire, the margin the driver keeps so it never runs dry and clicks, the same block again on the way out, and another millisecond in the D to A converter.

Now the arithmetic, because this is where the mismatch lives. A buffer is a count of samples, not a length of time. At 44.1 kHz, 64 samples is 1.45 milliseconds, 128 is 2.9, 256 is 5.8, 512 is 11.6 and 1024 is 23.2. Your software shows you the buffer, and 2.9 milliseconds looks like nothing. But the block happens twice, the converters take their cut at both ends, and the driver adds its margin on top. A competent USB interface at 128 samples typically measures eight to twelve milliseconds from sound in to sound out. That is the chain working correctly, and it is four times the number on the screen.

Buffer	At 44.1 kHz	What it is for
64 samples	1.45 ms	Tracking
128 samples	2.9 ms	Tracking
256 samples	5.8 ms	The setting that is never quite comfortable to track on and never quite stable to mix on
512 samples	11.6 ms	Mixing
1024 samples	23.2 ms	Mixing

Round Trip Against the Number Your Software Reports

Round trip latency is the only figure that describes what you are experiencing: a sound enters the interface, goes through the computer, and comes back out. Everything else is a component of it, including the figure your DAW displays, which is not a measurement. That is the driver declaring its own latency, and it is short for two ordinary reasons: it describes only the computer's portion of the path, so both converters are missing, and the margin the driver keeps in practice is often larger than the one it reports. It is answering a narrower question than the one you asked. Your DAW also uses that figure to line up overdubs, nudging each new recording earlier by the amount the driver claims, so when the claim is short every punch-in lands late by the size of the error.

So measure it. Patch output one into input one, record the loopback, and count the sample offset between what you sent and what came back. RTL Utility from Oblique Audio does this automatically on Mac and Windows and costs nothing. Run it at each buffer setting you use and write the numbers on tape stuck to the interface. Then play against it and trust the playing, which is the same order written down in how I listen: the number is the check on the account, never the frame around it.

What Ten Milliseconds Feels Like

A foot of air is about a millisecond. Sound covers roughly 1,125 feet per second, so ten milliseconds is standing eleven feet back from your amp. Nobody has ever called that a delay. So why does the same ten milliseconds in headphones ruin a take?

Because in a room everything moves together. The delay is identical at both ears, it is constant, and your body has had a lifetime to calibrate it. In headphones the delayed version arrives alongside an undelayed one: the acoustic sound of the instrument, the click of the pick, your own voice conducted through your skull. Two copies of the same event a few milliseconds apart, summed at your ears, is a comb filter. At three milliseconds the first notch sits around 167 Hz with another every 333 Hz above it, and that hollow, phasey quality is what people mean when they say the headphone mix sounds wrong.

The thresholds, plainly. Under about five milliseconds round trip nearly everyone plays normally. Between five and ten, sensitive players notice on percussive material and not much else. Past ten it stops being something you hear and becomes something you feel, which is when takes get stiff without anyone being able to say why. Past twenty you are no longer playing with the track, you are playing to a delay.

Why an Amp Simulator Feels Worse Than a Vocal at the Same Number

Put a singer and a guitarist through identical ten millisecond round trips and the guitarist complains first, louder, and with better reason.

A singer has a reference at zero latency that nothing can take away. You hear your own voice through bone conduction before any of it reaches a microphone, so your sense of when you are singing never drifts. What latency costs a singer is tone and comfort: the combing above, and the distraction of a voice in the cans that arrives after the one in your head. That will flatten a take, but the timing reference survives it.

An electric guitar through a simulator has none. The unamplified string is quiet and dull, so the only loud version of what you played comes back from the computer. The transient is percussive: a pick attack rises in a couple of milliseconds, and the ear resolves timing far more finely on a sharp attack than on a sustained vowel. The tactile cue is absolute, because your fingers know exactly when the pick left the string. So the guitarist is reconciling a hard physical event with a sound that audibly happened after it, with nothing else to hold onto. Struck is the worst case and sung the easiest, for that reason. A drummer playing a virtual instrument from pads is the hardest case in the building, and the one case where the usual fix does not apply at all.

Direct Monitoring, a Hardware Mixer, or Through the Computer

There are three monitoring paths, and choosing the wrong one is the most common cause of a latency problem that hardware was never going to solve.

Direct monitoring splits the input in analog and sends it to the headphones before it reaches a converter, usually as a blend knob between input and computer playback. Latency is a few microseconds, which is to say none. What it costs is processing: the performer hears a dry signal, and plenty of singers cannot perform against a bone dry voice in their own head. Many implementations also put the input dead center in both ears with little level control, which is fine for a guitar and unpleasant for a vocal.

A hardware DSP mixer is the same idea with those compromises removed. The interface runs a digital mixer on its own chip and you monitor through that instead of through the computer. RME's TotalMix FX, Universal Audio's Console, MOTU's CueMix and Antelope's onboard effects all work this way. Latency is the converter round trip and nothing else, usually under two milliseconds, and you get reverb, EQ and compression in the monitor path, plus preamp and amp models on the Universal Audio boxes. For most tracking this is the answer. Print dry, monitor wet, decide later is the oldest working method in the studio, and a DSP mixer is that method in a box.

Monitoring through the computer is necessary when the processed sound is the performance rather than a comfort. A virtual instrument played from a keyboard or pads has no analog signal to split, so there is nothing to monitor directly and the buffer is your only tool. The same holds when a player is phrasing against a specific amp simulator, with the gain and the delay shaping how they attack a note. Accept the computer path there, get the buffer down, and make the machine carry it.

The expensive mistake is reaching for a faster interface while a perfectly good direct monitoring switch sits unused on the front panel of the one you own. The hardware could have done the job, and the software was asked instead.

Monitoring path	Latency	What it costs you
Direct monitoring	A few microseconds, which is to say none	Processing. The performer hears a dry signal, and many implementations put the input dead center in both ears
A hardware DSP mixer	The converter round trip and nothing else, usually under two milliseconds	The price of the box. For most tracking this is the answer
Through the computer	The buffer, twice, plus both converters and the driver margin	Necessary when the processed sound is the performance rather than a comfort

Buffer Size Is Two Settings, Not One

Tracking and mixing want opposite things. While tracking there is a live performer and almost no plug-in load, so you spend processing headroom on timing: 64 or 128 samples. While mixing there is no live input and possibly forty plug-ins, so latency is free and stability is what you buy with it: 512 or 1024. Set 256 once and leave it, and you have bought yourself a tracking session that is never quite comfortable and a mix that is never quite stable, to save opening the preferences window twice a day. Open it. And if the machine cannot hold 128 while tracking, the fix is to freeze or bounce the heavy instrument tracks first, not to raise the buffer and hand the problem to the player.

Plug-in Delay Compensation, and the One Place It Bites

Some plug-ins cannot produce output until they have seen input they have not received yet. Linear phase EQ, look-ahead limiters, mastering compressors and most pitch correction all work this way. Each reports its delay to the host, the host delays every other track to match, and playback stays aligned. It is automatic, reliable, and while you are mixing it is invisible.

Here is where it bites. The host achieves that alignment by delaying everything else, and everything else includes whatever you are monitoring through. Put a linear phase EQ and a look-ahead limiter across the mix bus and every live input routed through the computer inherits their combined look-ahead. You can be running a 64 sample buffer, with your software reporting 1.5 milliseconds, and hearing yourself forty or fifty milliseconds late. The buffer control now does nothing, and that is the moment you start pricing a new interface.

The diagnostic takes ten seconds. Move the buffer from 64 to 512 and back. If how late you feel does not change, the delay is in a plug-in, and a new interface would have changed nothing. Fixes, in the order I would try them: monitor through the hardware and sidestep it entirely; use the mode your DAW already has, which is Low Latency Monitoring in Pro Tools, Low Latency Mode in Logic and Constrain Delay Compensation in Cubase; or take the look-ahead processors off the mix bus until tracking is done. The last one costs nothing and is what I do.

Sample Rate as a Latency Tool

Higher sample rates cost more than they return. Latency is the one exception, and the only reason I reach for 96 kHz on purpose. Because a buffer is a count of samples rather than a length of time, the clock rate decides what that count is worth. A 128 sample buffer is 2.9 milliseconds at 44.1 kHz and 1.33 milliseconds at 96 kHz. Every stage counted in samples shortens in proportion, the converter filters included, so raising the rate takes a real bite out of the round trip with no change of hardware. The transfers across the bus are fixed by the connection and do not shrink with it, so measure the new number rather than assuming half.

What it costs is everything else: double the file size, double the processing load, close to half the plug-in count. There is no sane way to change it partway through a project either, so it is a session-start decision rather than a knob to reach for when a take is going badly. An afternoon of a guitarist through a simulator is a good use of 96 kHz, because the feel improvement is real and the plug-in count while tracking is low anyway. A dense session running a lot of processing is not, so stay at 44.1 or 48 and solve the monitoring in hardware. The sample rate is a tool for the performance here, not for the fidelity.

When the Interface Actually Is the Problem

Sometimes it is. The driver is the real case, and it is why two interfaces with near identical specifications behave nothing alike. A well written driver holds a 64 sample buffer under load without a click; a poor one is unusable below 256, and that gap costs more milliseconds than any converter design ever will. It is a purchase criterion rather than a setting, and the dated version of that argument, with the download pages worth reading before the product pages, is in how to choose an audio interface.

The other three are shorter. No hardware monitoring path, or one nobody can work inside: mono only, no level control, no reverb. Processed sound at low latency the computer cannot deliver, which is what a Universal Audio Apollo, an RME running TotalMix or an Antelope actually sells you. Not cleaner converters, a monitor path with effects in it. And connection type, last and least: Thunderbolt gets round trip lower than USB, worth having once everything above is right and worth little before it.

One thing to rule out before blaming timing at all. If what you hear is thin, soft and dynamically flat rather than late, that is the headphone output running out of current, and it gets reported as the same complaint constantly. It depends on what is on your head more than on the box, which is what headphone impedance means and when to care.

The Short Version

Measure your real round trip with a loopback before changing anything, and write it on the interface. Under five milliseconds nobody notices, five to ten is workable for voice and hard for anything struck, over ten you feel it whether or not you can name it, and over twenty the take is gone.

Monitor through the hardware whenever there is an analog signal to split: direct monitoring if a dry sound is acceptable, the interface's own DSP mixer if the performer needs reverb to sing properly. Print dry either way. Go through the computer only when the processed sound is the part itself, and then run 64 or 128 samples with nothing look-ahead on any bus in the monitor path. Keep two buffer settings and use both, small to track and large to mix. If moving the buffer does not change how late you feel, stop moving it: the delay is plug-in delay compensation and the fix is on the mix bus. If all that is done and a guitarist still cannot play comfortably, take the session to 96 kHz and accept the halved plug-in count.

Buy a new interface for its driver, for a monitoring path you do not have, or for onboard DSP that lets a performer hear effects for nothing. Those are the three things that move this number. Buy one for a converter specification and you will install it, measure it, and find you spent a few hundred dollars to shave two milliseconds off a chain that was losing forty somewhere else. If it is genuinely one of the three, start at audio interfaces and take the rest of the decision from there.

---

Or Bring It In and Measure It

The listening room at Free River Studios, 215 E. Lewis St #301 in Livingston, is on the third floor, and it is a working studio as well as a room to listen in, so this is one of the few arguments in audio that gets settled in fifteen minutes. Bring your interface, your headphones and your laptop. We will loop an output back to an input, find what your round trip actually is at each buffer setting, and then you can play through it and find where your own threshold sits. It is by appointment, and you will leave knowing a number about your own rig that no specification sheet was going to give you.