44.1 vs 48 vs 96 kHz: Which Sample Rate Should You Record At?

This setting gets argued about far out of proportion to what it changes. The honest position is that 48 kHz is right for nearly every home studio project, that 44.1 kHz is a delivery format rather than a working decision, and that 96 kHz solves two specific problems and doubles your storage and processing cost to do it. What follows is the arithmetic that supports each of those, including the file size figures you can check yourself.

Any current interface handles all three rates, so this is a setting rather than a purchase. Set it once per project, before you record the first thing, and do not change it mid-session.

What does the sample rate actually determine?

One thing, directly: bandwidth. The Nyquist theorem says a sampled system can represent frequencies up to exactly half the sample rate, and nothing above it. So the ceiling on your recording is fixed the moment you choose the rate.

Work the headline row through by hand, because it is the number worth being able to reproduce. At 48 kHz, 24 bits is three bytes, two channels doubles it: 48,000 x 3 x 2 = 288,000 bytes per second, or 288 kB/s. Multiply by 60 and you get 17,280,000 bytes in a minute. Your operating system divides that by 1,048,576 rather than by a million, which is why the file shows up as about 16.48 MB. Both figures are correct, they just use different definitions of a megabyte, and that is the source of most of the confusion when people compare notes.

A human ear at its best reaches roughly 20 kHz, and that ceiling drops steadily with age, so even 44.1 kHz already captures more bandwidth than any listener will hear. Nobody chooses 96 kHz because they want the 30 kHz content. The arguments are all indirect, and they are below.

Why do 44.1 and 48 both exist?

History, not engineering preference. 44.1 kHz came out of storing digital audio on video tape in the late 1970s: the number falls out of the line and field rates of the video formats that early PCM adaptors piggybacked on, and it then became the CD standard and stayed. 48 kHz came from professional video and broadcast, where a rate that divides evenly against film and video frame rates makes editing and synchronisation far cleaner.

Today that split maps almost perfectly onto the delivery question. Anything going to video, streaming platforms, games or broadcast wants 48 kHz. A CD master wants 44.1 kHz. Streaming services accept both and convert, and modern sample rate conversion is transparent enough that this is no longer a reason to lose sleep, but converting is still one process you can avoid by picking the right rate at the start.

The 8.8 percent storage difference between them is not a real consideration on any modern drive. Pick by deliverable.

What does the anti-alias filter have to do?

Any frequency above the Nyquist limit that reaches the converter does not simply disappear. It folds back down into the audible range as an inharmonic tone, which sounds like a metallic whistle riding on top of the material. Preventing that is the job of the anti-alias filter, and its difficulty depends directly on the sample rate.

• At 44.1 kHz the filter must pass 20 kHz and be effectively silent by 22.05 kHz. That is a 2.05 kHz transition band, which is a very steep slope.
• At 48 kHz the same filter has until 24 kHz, so the transition band is 4 kHz, roughly twice the room.
• At 96 kHz the filter has until 48 kHz and the design becomes trivial.

In 1985 that difference was audible, because those filters were analogue and steep analogue filters ring and shift phase near the corner. It is much less of an issue now. Modern converters are delta-sigma designs that sample internally at a very high rate and do the steep filtering in the digital domain, so the analogue filter can be gentle regardless of the rate you selected. What remains is a choice of digital filter type, and that is where the small audible differences between converters live: minimum phase filters shift phase near the top but do not pre-ring, linear phase filters keep phase intact and produce a small pre-echo ahead of sharp transients. Some interfaces let you choose. It is a far smaller effect than microphone position.

What does a higher rate cost in storage and CPU?

Both scale linearly, which makes the budgeting easy and the consequences obvious once you write them down for a real session rather than for a single track.

Take 24 mono tracks of a 4 minute song, recorded at 24 bit. At 48 kHz each track is 8.64 million bytes per minute, so the session's audio comes to roughly 829 million bytes. At 96 kHz the same session is 1659 million bytes, comfortably over a gigabyte and a half. Then remember that real sessions carry alternate takes, comps and playlists, so multiply by three or four for how the folder actually ends up.

CPU behaves the same way. A plugin at 96 kHz has twice as many samples to process each second as the same plugin at 48 kHz, so as a rule of thumb your plugin count halves. Convolution reverbs, linear phase equalisers and anything with oversampling of its own are the first things to fall over. Track counts, disk streaming and automation all scale similarly.

Does a higher rate reduce latency?

Yes, and this is the one genuine performance benefit. Buffer latency is buffer divided by sample rate, so at a fixed buffer size a faster rate empties it sooner.

Read the table carefully, though, because the obvious conclusion is the wrong one. A 128 sample buffer at 96 kHz gives 1.333 ms. A 64 sample buffer at 48 kHz gives 1.333 ms, which is the same number, while processing half as many samples per second. Doubling the rate to chase latency is paying twice the CPU for something a buffer setting already gives you. The only wrinkle is that very small buffers mean more driver callbacks per second, and some systems cope better with a larger buffer at a higher rate. Test both on your machine rather than assuming. The latency buffer calculator gives you the figures for any combination and the sample rate and latency chart has the full grid. If latency is the problem you came to solve, start with how to fix latency in your DAW, because it is usually a plugin rather than a rate.

What about plugin aliasing and oversampling?

Here is the strongest technical argument for 96 kHz, and also the reason it has mostly stopped applying.

Any nonlinear process generates new harmonics that were not in the input. Distortion, saturation, tube and tape emulation, clipping, and compressors with fast nonlinear behaviour all do this. Those new harmonics can easily land above your Nyquist limit, and when they do they fold back into the audible band as inharmonic content that has no musical relationship to the source. It sounds like a metallic edge or a harshness that gets worse the harder you drive the plugin.

Running the whole session at 96 kHz raises the fold-back point to 48 kHz, so the aliasing products land where you cannot hear them. That works. It is also a very inefficient way to fix a problem that only exists inside a handful of plugins, since you are paying double the processing cost on all one hundred tracks to fix three of them.

The efficient fix is plugin-level oversampling: the plugin internally upsamples to 2x, 4x, 8x or 16x, does its nonlinear work up there, filters, and comes back down. Most serious modern saturation and distortion plugins do this, and many expose it as a quality setting you can turn up for the mixdown and down while writing. If you own a plugin that aliases badly and offers no oversampling, the honest answer is to replace the plugin rather than to double the session's sample rate.

When is 96 kHz genuinely the right choice?

Three cases, and they are narrow but real.

1. Heavy downward pitch manipulation. If you are going to pitch a recording down an octave, everything above 10 kHz in the original ends up below 5 kHz in the result. Content that was ultrasonic in the original becomes audible detail rather than absence. Sound designers recording metal, machinery, animals and impacts for later manipulation record high for exactly this reason.
2. Time stretching by large factors. The same logic. More source information means the algorithm has more to work with, and the artefacts are less exposed.
3. A specific plugin or hardware path that requires it. Some analogue modelling designs are documented as behaving differently at higher rates, and some measurement and calibration work needs bandwidth above 24 kHz.

Notice what is not on that list: recording a band, mixing a song, tracking a podcast, or anything to do with how good the finished record sounds. If you are recording music to release, 48 kHz is not a compromise you are settling for.

What about bit depth?

Bit depth is the more consequential setting and it is barely argued about, which is backwards. Sample rate sets bandwidth; bit depth sets dynamic range, which is the axis your gain knob lives on.

Record at 24 bit, always. It gives roughly 144 dB of theoretical dynamic range, which is more than any microphone, preamp or room can deliver. That is why the site's whole gain-staging position works: aim for around -18 dBFS RMS with peaks near -10 to -6 dBFS, leave enormous headroom, and stop chasing a hot signal. There is no penalty for recording conservatively at 24 bit, and there is a large penalty for clipping the converter. See how to set recording levels for the routine.

A handful of interfaces, such as the Zoom UAC-232 (32-bit float) , record in 32-bit float, which removes the input gain decision almost entirely because the format can represent levels far beyond digital full scale. It costs a third more storage than 24 bit and it only helps if your DAW handles the format properly. It is a genuine workflow change rather than a marketing spec, and it does nothing for a badly placed microphone.

So what should you set?

Two rules that prevent the common mistakes. Set the rate before you record anything, because changing it partway through a project means converting files and creates a source of subtle sync and pitch errors. And set every device in the chain to the same rate: your interface, your DAW project and any digital device connected over ADAT or S/PDIF. A clock mismatch produces clicks or a project that plays back at the wrong pitch, which is one of the most commonly misdiagnosed faults in a home studio. The signal flow guide covers where the clock lives, and the interface database lists the maximum rate for every model if you are shopping.

Related comparisons and tools

• USB vs Thunderbolt interfaces: the bandwidth that a higher sample rate consumes
• Audio interface vs USB microphone: where the converter actually lives in your chain
• Latency buffer calculator: your exact latency at any rate and buffer
• Sample rate and latency chart: the full grid without typing anything
• MOTU M2 review: the only real front panel meter at the price