Audio Interface Latency Explained: Buffer Size, Round-Trip Latency and Direct Monitoring

A 32-sample buffer is a setting, not a promise about what reaches a musician’s ears. The useful question is how long the entire monitoring path takes, whether that path includes the sound the performer needs, and whether it stays reliable through a real session. We would buy a predictable interface at a stable buffer before buying a smaller number on a control panel.

Start by naming the path

For a microphone monitored through a DAW, the route is analog input, A/D conversion, input driver and buffers, DAW processing, output buffers and driver, D/A conversion, then the headphone amplifier. Acoustic propagation and a digital headphone system can add further delay. A virtual instrument follows a different path: controller scanning and MIDI delivery precede instrument processing and the output path, but there is no microphone A/D stage. Do not compare those experiences as if they were one latency specification.

Recording placement is a third question. A DAW can move a recorded region according to reported device delays so that it lands near the intended timeline position. That correction does not remove the delay the performer heard while recording. Good-looking waveform alignment is not evidence of an instantaneous monitor path.

The arithmetic that a buffer label leaves out

A block of B samples at sample rate Fs lasts 1,000 × B/Fs milliseconds. The following values are calculations, not measurements of an interface. The last column merely adds one input block and one output block of equal size; real implementations may buffer differently.

Sample rate	Block	One block	Two-block arithmetic
48 kHz	64 samples	1.33 ms	2.67 ms
48 kHz	128 samples	2.67 ms	5.33 ms
96 kHz	64 samples	0.67 ms	1.33 ms

Converter filters, safety offsets, transport scheduling and processing delay still have to be accounted for. Doubling the sample rate shortens a fixed sample-count buffer but increases the samples that must be processed per second. It can also reduce available digital I/O. That trade belongs in the session plan, not in a blanket recommendation to record everything at 96 kHz.

Why a fast computer can still crackle

Audio has recurring deadlines. A live input with a long serial effects chain cannot simply spread every dependent calculation across every idle core. A driver interruption, thermal limitation or demanding processor can miss the next deadline while an operating-system CPU meter looks comfortable. A system that plays an already rendered mix is not necessarily capable of playing a virtual piano through the same session at a tiny buffer.

Use the DAW’s real-time performance display as well as total CPU load. Compare playback with record-armed monitoring, then remove the most expensive processors one at a time. Freezing an unrelated instrument may help overall capacity; it cannot eliminate the processing delay of a lookahead limiter still sitting on the live cue bus. Our CPU, memory and storage guide explains how to distinguish these bottlenecks.

Direct monitoring trades one problem for a routing decision

An analog monitor path bypasses the computer. An interface’s digital mixer usually still includes conversion and local processing, but avoids the host round trip. Calling both “zero latency” hides that distinction. Either can be an excellent way to give a singer a responsive cue.

What matters is what the performer hears. A guitarist depending on a native amp simulator needs the DAW path; a singer may accept a dry direct signal plus locally generated reverb. DSP effects are specific to the interface and its software. Do not assume an onboard mixer can run arbitrary native plug-ins.

Disable the unintended duplicate path. Hearing direct input and delayed DAW input together can create comb filtering, a slap or misleading tonal changes. Keep the backing tracks in the cue mix, but establish deliberately which route carries the live source. In loopback tests, turn off any route that feeds the returned input back into its own output.

Measure interface delay and session delay separately

Oblique Audio’s RTL Utility measures a physical output-to-input loopback. Its documentation distinguishes that interface/driver measurement from externally measuring an analog-input-through-DAW monitoring path. Use the supplied procedure and log the driver, firmware, operating system, sample rate and buffer. This is a useful baseline, not a measurement of your master-bus plug-ins.

1. Connect a line output to a suitable line input with monitors muted, low initial level and phantom power off on the test input. Do not connect a power-amplifier output.
2. Select the intended driver and channels. Disable direct feedback and avoid automatic gain changes. Confirm the return is clean and not clipped.
3. Measure several times at each working buffer. Investigate inconsistent results instead of keeping only the best result.
4. Open a representative DAW session and test sustained monitoring, punching in, cue mixes and virtual instruments. Check whether any low-latency mode bypasses sends or effects.
5. For the actual monitored analog-through-DAW delay, use a suitable external reference/capture method or documented analyzer procedure. If using timeline offsets, account explicitly for the DAW’s recording compensation.

Report baseline RTL and loaded-session stability independently. Neither a manufacturer’s display nor a brief dropout-free test establishes reliability for all workloads.

Delay compensation cannot predict a live performance

During playback, the DAW can delay faster paths to align them with slower ones. It cannot make an unperformed note arrive early. Linear-phase EQ, lookahead dynamics and some oversampling modes add algorithmic delay; disabling a processor’s audible effect may not remove its reported latency. Check the host’s actual bypass or low-latency behavior.

For tracking, save a cue configuration with high-latency processing removed from relevant buses. For mixing, increase the buffer and restore the intended processing. Apple’s explanation of plug-in latency compensation describes the underlying distinction; menu names and available modes depend on the current host version.

Buy a monitoring workflow

There is no universal millisecond threshold that makes every vocalist, drummer and keyboard player comfortable. Their direct acoustic sound, instrument attack, headphones and expectations differ. Ask performers to evaluate the actual cue path, at sensible levels, rather than declaring a number “inaudible.”

We would compare RME’s Babyface Pro FS documentation and other candidate manuals for routing and driver support, then verify the exact system before the return deadline. That is a research example, not a TAM performance ranking. Read choosing an interface for the I/O and cue-mix questions that belong beside latency. A small buffer that loses a take has failed the only comparison that matters.

Engineering explainer and proposed evaluation procedure; TAM did not perform new interface measurements for this article.