Sample Rate and Bit Depth, Without the Myths

Digital audio is described by two numbers that are constantly confused with each other. Sample rate is measured in kHz, bit depth in bits, and the folk explanation — that both are about "resolution," so higher must mean closer to the original waveform — is wrong about what each one does. They control two entirely separate properties, and understanding which is which explains most of the practical decisions.

Sample rate sets a frequency ceiling, not smoothness

A sample rate of 44,100 Hz means the signal's amplitude was measured 44,100 times per second. The intuitive reading is that this produces a jagged staircase approximating a smooth curve, and that a higher rate would produce a finer staircase closer to the truth.

That picture is wrong, and the Nyquist–Shannon sampling theorem says why. If a signal contains no frequencies at or above half the sample rate, the samples don't approximate it — they determine it exactly. There is precisely one band-limited waveform passing through any given set of samples, and reconstruction recovers it perfectly. The staircase is an artifact of how the samples are drawn on a screen, not of what they represent. A properly designed converter doesn't output steps; it outputs the unique smooth curve those points imply.

So sample rate buys exactly one thing: a ceiling. At 44.1 kHz the ceiling — the Nyquist frequency — is 22.05 kHz, comfortably above the roughly 20 kHz upper limit of human hearing, which is why CD audio settled there. At 48 kHz, standard in video, the ceiling is 24 kHz.

Frequencies above the ceiling are not merely lost. They alias: a 25 kHz tone recorded at 44.1 kHz folds back down and reappears as an audible 19.1 kHz tone that was never played. Because of this every converter applies an anti-aliasing low-pass filter before sampling. That filter has to be steep, and steep filters have side effects near the cutoff — which is the honest argument for 96 kHz recording. It's not that you capture more audible content; it's that the filter can be gentler because it has an inaudible octave to roll off in. Once you've made a finished recording, converting it upward gains nothing at all. The information above 22 kHz was never captured, and no resampling invents it.

Bit depth sets the noise floor

Bit depth is a different axis entirely. Each sample's amplitude must be stored as a number with finite precision, and rounding to the nearest representable value introduces a small error. Across many samples, that rounding error behaves like a faint layer of noise underneath the signal.

The relationship is close to exactly 6.02 dB of dynamic range per bit. 16-bit gives about 96 dB between the loudest representable signal and the noise floor. 24-bit gives about 144 dB, which is below the self-noise of any microphone and any room — you are recording the noise of your equipment long before you reach it.

So bit depth doesn't make the waveform "smoother" either. A 16-bit and a 24-bit recording of the same loud passage are nearly identical; the difference lives entirely in how quiet a sound can get before it disappears into quantization noise.

The reason 24-bit matters is headroom during recording. You must set levels so the loudest unexpected peak doesn't clip, which means the average signal sits well below maximum. With 16 bits, recording 20 dB down leaves you about 76 dB of usable range and any later gain lifts the noise floor with the signal. With 24 bits you can record conservatively and still have far more range than you need. For a finished file that will only be played back, 16-bit is genuinely transparent, which is why the format has outlasted every attempt to replace it for distribution.

There's one refinement. Reducing bit depth at the end — 24-bit master down to 16-bit — should be done with dither: a tiny amount of deliberately added noise before rounding. Without it, quantization error correlates with the signal and produces distortion that sounds like a gritty edge on quiet fades. With it, the error is decorrelated and becomes plain hiss, which the ear tolerates far better at the same level. Trading a distortion you notice for a noise you don't is one of the better bargains in signal processing.

Where the bytes actually go

Uncompressed PCM is easy to compute: sample rate × bit depth × channels. CD audio is 44,100 × 16 × 2 = 1,411,200 bits per second, about 10 MB per minute.

Lossy formats reach roughly a tenth of that by discarding what a psychoacoustic model predicts you can't hear — the subject of what MP3 actually throws away. What matters here is that in a lossy file, bitrate is the parameter that governs quality, and the sample rate and bit depth of the source barely matter. Encoding a 24-bit 96 kHz master to a 128 kbps MP3 gives you a 128 kbps MP3. The encoder resamples and requantizes internally, and the extra source precision is discarded in the first step.

This is the single most common mistake when converting audio: raising the sample rate hoping to improve a lossy file. It increases the file size and changes nothing audible, because the ceiling was set when the original was encoded and cannot be raised afterward.

Practical settings

• Recording: 24-bit, 48 kHz. The bit depth is where the real safety margin is. Higher sample rates are a legitimate choice if you'll do heavy pitch-shifting or time-stretching — those operations move content around in frequency and can pull ultrasonic material into the audible band — but the benefit is in processing, not capture.
• Distribution: 16-bit, 44.1 or 48 kHz, dithered on the way down. This is transparent and has been for forty years.
• Speech: mono, and often a lower sample rate. Intelligible speech lives mostly below 8 kHz; a podcast at 64 kbps mono is a reasonable trade, and halving the channel count halves the data before any codec is involved. When shrinking an audio file, dropping stereo to mono for a single speaker is almost always a bigger win than shaving the bitrate further.
• Never upsample a finished file. It cannot add information, and the larger file will convince someone downstream that it's higher quality than it is.

The two numbers answer two different questions. Sample rate: how high a frequency can this file contain? Bit depth: how quiet a sound can this file contain? Almost every audio myth comes from collapsing those into a single vague idea of resolution.