Sample Rate and Bit Depth: What They Do, and the One Case Where It Bites
Data engineer who loves building high-performance data and web-related tools. Creator of SlowedReverbMaker.net, implementing browser-side digital signal processing (DSP) to democratize audio editing.
Sample rate and bit depth are the two numbers that define a piece of digital audio, and they are the subject of more confident nonsense than any other topic in the field.
The short version is that they control two completely different things, that the standard values are genuinely sufficient for listening, and that the endless arguments about higher numbers are mostly about recording and production rather than playback.
But there is one situation where sample rate stops being an abstraction and starts audibly damaging your work — and it is a situation anyone making sped-up edits walks into without warning. That is at the end, with a measurement.
Try this while you read: Nightcore Maker
Speed up songs and raise pitch for nightcore edits. Free, in your browser, no signup.
1. The Two Numbers, Defined
Sample rate and bit depth control completely different things, and most of the confusion around them comes from treating them as two dials on the same quality knob.
- Sample rate
- How many times per second the waveform's position is measured. 44,100 for CD audio, 48,000 for most video work. It determines the highest frequency the recording can contain.
- Bit depth
- How precisely each individual measurement is described. 16 bits gives 65,536 possible values per sample; 24 bits gives about 16.7 million. It determines dynamic range, not frequency.
- Nyquist limit
- Half the sample rate — the highest frequency a recording can represent. At 44.1 kHz that is just under 22.05 kHz.
- Aliasing
- What happens when a frequency above the Nyquist limit is asked to exist. It cannot, so it folds back and reappears as a lower frequency, mirrored around the ceiling.
2. Sample Rate Is About Frequency
The Nyquist–Shannon sampling theorem gives the rule: to represent a frequency, you must sample at more than twice that frequency.
Why 44.1 kHz Specifically
Human hearing tops out somewhere around 20 kHz in young people with undamaged ears, and declines with age — most adults do not hear much above 16 kHz.
44.1 kHz places the limit just above the range anyone can perceive, with a small margin for the filtering required at the boundary. It was not an arbitrary choice.
The Stair-Step Picture Is Wrong
A common claim is that higher sample rates produce a 'smoother' waveform, usually illustrated with a jagged stair-step diagram.
That picture is simply incorrect. Reconstruction does not connect samples with straight lines. The sampling theorem guarantees that for a band-limited signal there is exactly one continuous waveform passing through the sample points, and the converter reproduces it. There is no jaggedness to smooth out.
3. Bit Depth Is About Dynamic Range
Bit depth is a separate axis entirely, and it does not control 'detail' in any frequency sense.
What the Bits Actually Buy
What bit depth controls is the distance between the loudest representable signal and the noise floor created by rounding each measurement to the nearest available value. Roughly 6 dB of dynamic range per bit — which puts 16-bit at about 96 dB and 24-bit at around 144 dB.
96 dB is a very large range. It comfortably exceeds the difference between a quiet room and a level loud enough to be uncomfortable, which is why 16-bit is entirely adequate for delivery.
Where Extra Bits Genuinely Help
During recording and processing, not playback. When tracking, you leave headroom so an unexpected loud moment does not clip — which means recording at a lower average level and using less of the available range.
24-bit gives you room to be conservative without pushing quiet material toward the noise floor. It is insurance for the production process, not a better listening experience.
4. So Do Higher Numbers Help?
Three different answers depending on what you are doing, and it is worth separating them.
- For listening: no, with reasonable confidence. Content above 20 kHz is inaudible and 96 dB of dynamic range exceeds what any normal playback situation uses. A 192 kHz file is over four times the size of a 44.1 kHz one and contains a great deal no human will perceive.
- For recording and mixing: 24-bit is worth it for the headroom reasons above, and is essentially free. Higher sample rates have a narrower case — some processing benefits from room above the audible band — but the benefits are modest and the file sizes are not.
- For editing existing audio: your source already has a sample rate and bit depth, and nothing you do afterwards adds information that was never captured. Converting a 44.1 kHz file to 96 kHz produces a larger file containing the same music.
5. The Case Where It Bites: Aliasing
Here is where sample rate stops being theoretical, and it is specific to something many people do casually.
Speeding Up Pushes Content Toward the Ceiling
Varispeed — the technique behind nightcore edits, sped-up remixes, and every slowed edit — works by reading the samples at a different rate. Speed the track up and every frequency multiplies by the same factor.
But the ceiling does not move. If the file is 44.1 kHz, nothing above 22.05 kHz can be represented, no matter what you do. So what happens to a 15 kHz cymbal shimmer when you play the track at double speed and it tries to become 30 kHz?
It cannot go there. In a system that does not filter first, it folds back — reappearing as a lower frequency, mirrored around the ceiling. Unlike most audio problems this does not sound like a degraded version of the original. It sounds like content that was never in the recording, at frequencies with no musical relationship to anything else.
An Important Caveat on Scope
This measurement was taken with a Node implementation of the Web Audio API, not in a browser. Implementations differ in whether and how they filter before resampling, so a given browser may handle the same operation better.
The underlying principle is not implementation-specific — resampling without band-limiting aliases, always — but the severity you encounter in practice may not match the figures above.
What It Sounds Like in Real Music
Real music does not consist of single tones, so aliasing from a speed-up shows up as a general grittiness or metallic edge in the top end rather than an identifiable pitch. It is worst on material with a lot of high-frequency energy: cymbals, hi-hats, sibilance, distorted guitars.
It also gets worse the faster you go, which means extreme nightcore settings are exactly where to listen for it. If the highs sound harsh in a way that adjusting treble does not fix, this is a candidate explanation — and starting from a higher-sample-rate source raises the ceiling and gives the shifted content somewhere legitimate to go.
6. And Why Slowing Down Is Safe
The mirror image is worth stating, because it is one of the few places where slowed edits have an inherent advantage.
Slowing a track moves every frequency downward. A 15 kHz shimmer at 0.85x becomes 12.75 kHz, comfortably inside the representable range. Nothing approaches the ceiling, nothing folds back, and no aliasing occurs — the operation is safe by construction, no matter how far you take it.
The Short Version
Sample rate sets the highest frequency a recording can hold; bit depth sets the distance between the loudest signal and the noise floor. 44.1 kHz and 16 bits are sufficient for listening, higher bit depth is genuinely useful while recording, and higher sample rates matter far less than the marketing suggests.
The exception is speeding audio up, where content can be pushed past the limit and fold back as frequencies that were never in the music. Slowing down has no equivalent failure. If you take one practical thing from this: use the best source you have, especially for fast edits, because that is the one variable you control that nothing downstream can repair.
Method
The aliasing figures were measured by rendering a synthesised 15 kHz sine at amplitude 0.8 through this site's `renderProcessed` at 1.0x, 1.5x, and 2.0x, then measuring energy at specific frequencies with a Goertzel filter over a one-second window taken from the middle of the output. The environment was Node with the `node-web-audio-api` implementation of the Web Audio API, which is not necessarily identical to any browser's resampler — the caveat in section 4b is there for that reason. The predicted fold-back frequencies (21.6 kHz and 14.1 kHz) were calculated as |sample rate − intended frequency| before measuring, not fitted afterwards.