Audio latency is one of those subjects you think you have nailed until the day the vocal comes back in the headphones slightly late, a plugin nudges the kick off the grid, or the computer starts crackling mid-take. Behind a single word hide several distinct phenomena. Let us sort them out: buffer size, round-trip latency, the hidden sources and the concrete fixes so you never have to guess on this ground again.
What latency actually is
Latency is the time between the moment a signal enters your digital system and the moment it leaves it. A computer does not process audio sample by sample in real time: it works in packets. It gathers a small block of samples in a memory buffer, processes it in one pass, then moves to the next. This block-based approach is what makes computer-based production possible; it is also the main source of latency.
Latency is measured in milliseconds, and the threshold that matters is perception. Below 5 to 10 ms, a musician monitoring live feels almost nothing. Around 10 to 20 ms, the discomfort becomes noticeable for a singer or a drummer. Beyond that, playing in time gets uncomfortable, then impossible. For scale: sound travels about 3.4 metres through air in 10 ms. In other words, 10 ms of latency is like standing three and a half metres from your amp – nothing dramatic in itself, but it adds up fast.
Buffer size: the core trade-off
Buffer size is set in your software or your interface control panel, in samples: 32, 64, 128, 256, 512, 1024. The smaller the buffer, the lower the latency, but the harder the CPU works and the higher the risk of dropouts. The larger it is, the higher the latency, but the more the system breathes.
Converting to milliseconds depends on the sample rate. Here are the per-direction values at 48 kHz:
| Buffer size | Latency at 48 kHz (one direction) |
|---|---|
| 32 samples | 0.7 ms |
| 64 samples | 1.3 ms |
| 128 samples | 2.7 ms |
| 256 samples | 5.3 ms |
| 512 samples | 10.7 ms |
| 1024 samples | 21.3 ms |
The field rule is simple: small buffer to record (the performer must hear themselves without lag), large buffer to mix (you are loading dozens of plugins and no longer need to play live). Raising the sample rate slightly lowers latency at a given buffer size, but adds just as much processing load: it is not the right lever for responsiveness.
Round-trip latency: the number that really matters
The figure your buffer shows tells only half the story. What the musician actually feels is round-trip latency: the time it takes the signal to enter (analogue-to-digital conversion), pass through the input buffer, be processed, go through the output buffer, and come out again (digital-to-analogue conversion). In practice, round-trip latency is well above double the buffer value, because converter and driver latency add to it.
That is why two interfaces set to the same buffer can feel very different: driver and converter quality make all the difference. A good interface actually quotes its real round-trip latency, not just the buffer size. The stability of the digital clock and the quality of conversion, which we cover elsewhere, are part of that chain.
The other sources of latency
The buffer is not the only culprit. Several elements add their own delay, often without the user knowing:
- A/D and D/A converters: each conversion stage introduces a fixed delay, usually under a millisecond, but real.
- Drivers: ASIO on Windows, Core Audio on macOS. A poorly written driver can add several milliseconds; it is one of the genuine criteria when choosing an interface.
- Some plugins: look-ahead processing (limiters, predictive de-essers), linear-phase EQs and oversampling introduce sometimes significant latency. The software compensates automatically at mixdown – that is plugin delay compensation (PDC) – but that compensation is incompatible with live playing.
- Wireless and Bluetooth: to be avoided entirely for any real-time monitoring.
Direct monitoring: bypassing the problem
The most effective solution is not to reduce latency but to short-circuit it. Direct monitoring (or hardware monitoring) sends the input signal to the performer’s headphones straight inside the interface, without going through the computer. Latency then drops to an inaudible value, whatever buffer size you chose for the computer.
Most interfaces offer a low-latency internal mixer that blends the input signal with the return from the computer: the performer hears themselves live over a playback that can run with a large buffer. The only limit: the musician does not hear effects computed by the computer (reverb, autotune). The most refined interfaces therefore build in a hardware reverb or onboard DSP to make the vocal comfortable again without reintroducing latency. Choosing a well-designed audio interface largely comes down to this.
Setting your latency in practice
A few habits are enough to stop suffering latency:
- When recording, set a small buffer (64 or 128) and enable direct monitoring as soon as a performer plays in real time.
- When mixing, raise the buffer (512 or 1024): you get processing power back and latency no longer matters since nobody is playing.
- Freeze or bounce hungry virtual instruments to audio: you free the CPU and stabilise the system.
- Keep an eye on gain structure in parallel: a healthy chain starts with clean levels. Our guide to gain staging in the digital domain usefully rounds this out, as do our notes on the mic preamp and on choosing studio monitors.
My take
Latency is not a flaw to fight at all costs, it is a parameter to manage. The habit that changes everything fits in one line: small buffer and direct monitoring to capture, large buffer to mix. Once that move is automatic, the question disappears and you get back to what matters – playing without thinking about it. The real difference between two interfaces is not on the marketing sheet but in the real round-trip latency and the robustness of the drivers. That is where to look before you buy.
