Farid Zaripov
Two numbers describe every uncompressed digital audio file: how often the signal is measured, and how precisely each measurement is stored. Sample rate and bit depth are simple ideas surrounded by a remarkable amount of folklore — most of it involving the belief that higher is always better. Higher is sometimes better, sometimes irrelevant, and occasionally counter-productive.
Sample Rate: How Often the Signal Is Measured
Sample rate is the number of measurements taken per second, in hertz. At 44 100 Hz, the waveform is sampled 44 100 times every second.
The number that matters is set by the Nyquist–Shannon theorem: a sample rate of f can perfectly represent frequencies up to f/2. So 44.1 kHz captures everything up to 22.05 kHz, and 48 kHz reaches 24 kHz. Human hearing tops out around 20 kHz in young adults and declines with age — by 40 many people cannot hear above 16 kHz. That is precisely why 44.1 kHz was chosen for the compact disc: it covers the audible range with a small margin for the anti-aliasing filter to roll off.
| Sample rate | Captures up to | Where it is used |
|---|---|---|
| 8 kHz | 4 kHz | Telephone speech |
| 16 kHz | 8 kHz | Wideband voice, speech recognition |
| 44.1 kHz | 22.05 kHz | CD, music streaming, most released music |
| 48 kHz | 24 kHz | Video, broadcast, most audio interfaces by default |
| 96 kHz | 48 kHz | Studio recording, heavy pitch or time manipulation |
| 192 kHz | 96 kHz | Archival work, scientific measurement |
Does 96 kHz Sound Better?
For playback of finished material, essentially no — the extra bandwidth sits entirely above human hearing. Where higher rates genuinely help is during production:
- Extreme pitch shifting and time stretching. Slowing a 96 kHz recording to half speed brings content from 48 kHz down to 24 kHz — still above hearing but with room for the processing artifacts to fall outside the audible band.
- Non-linear processing. Saturation, distortion and some compressors generate harmonics above the original content. At a higher rate those harmonics have somewhere to go instead of folding back into the audible range as aliasing.
- Latency. Buffer latency scales with sample rate: a 128-sample buffer is 2.9 ms at 44.1 kHz and 1.3 ms at 96 kHz. That matters when monitoring a live performance through software effects.
The costs are real: double the rate is double the file size and roughly double the CPU load, and 192 kHz can actually degrade quality because some analog-to-digital converters perform worse at their maximum rate, and some plugins introduce more intermodulation distortion when fed ultrasonic content.
Pick the Rate That Matches the Destination
The one thing genuinely worth avoiding is unnecessary conversion. 44.1 and 48 kHz are not simple multiples of each other, so converting between them requires interpolation, and poor sample-rate conversion is audible in a way that the original choice never was.
- Audio destined for video: record and deliver at 48 kHz. Every video standard uses it.
- Audio destined for music streaming or CD: 44.1 kHz.
- Unsure: 48 kHz is the safer default, since video is the more common destination and streaming platforms accept it.
Bit Depth: How Precisely Each Sample Is Stored
Bit depth sets how many discrete levels each measurement can take. 16-bit gives 65 536 levels, 24-bit gives 16.7 million. In practice, bit depth translates directly into dynamic range — the distance between the loudest signal and the noise floor:
Dynamic range ≈ 6.02 dB × number of bits
| Bit depth | Dynamic range | Notes |
|---|---|---|
| 8-bit | 48 dB | Audibly noisy; retro game audio |
| 16-bit | 96 dB | CD standard; exceeds most listening environments |
| 24-bit | 144 dB theoretical | Recording standard; real converters reach 120 dB |
| 32-bit float | Effectively unlimited | Clipping is recoverable; used inside DAWs and some recorders |
Note the gap between theory and practice at 24-bit. No analog circuit is quiet enough to deliver 144 dB — thermal noise in the electronics puts a ceiling around 120–125 dB on the best converters. The extra bits are still useful, but as working headroom rather than as audible resolution.
Why Record at 24-bit Then?
Because of headroom, not because 16-bit sounds bad. When recording, you must leave space for unexpected peaks — someone laughs, a guitarist digs in — so you set levels to peak around −12 to −18 dBFS. At 16-bit, that discards 12 to 18 dB of your available range and pushes the quiet passages closer to the noise floor. At 24-bit, the noise floor is so far down that conservative levels cost nothing audible.
This is the practical rule that follows: record at 24-bit, deliver at 16-bit. The extra precision protects you during capture and editing; the final file does not need it.
32-bit Float
Rather than fixed steps, 32-bit float stores each sample as a floating-point number, which means there is no hard ceiling at 0 dBFS. A recording that clips can be pulled back down afterwards with the waveform intact. Field recorders that offer it effectively remove the risk of a ruined take. Inside a DAW, all processing happens in 32-bit float or higher regardless of your file format — which is why a chain of plugins does not accumulate quantisation noise.
Dither: The Step People Skip
Reducing bit depth — 24-bit master down to a 16-bit file — means rounding every sample to a coarser grid. Done naively, the rounding error correlates with the signal and produces distortion that is most audible on quiet fades, where it sounds like a gritty crackle rather than clean silence.
Dither adds a tiny amount of carefully shaped noise before truncation, which decorrelates the error and converts distortion into a barely perceptible hiss roughly 90 dB down. The rules are short:
- Apply dither only once, as the very last step before writing the final reduced-depth file.
- Do not dither when staying at the same bit depth, or when exporting to 24-bit or 32-bit float.
- Do not dither repeatedly through a chain — each pass adds its own noise.
File Size: The Arithmetic
For uncompressed PCM audio:
Bitrate (kbps) = sample rate × bit depth × channels ÷ 1000
| Format | Bitrate | Per minute | Per hour |
|---|---|---|---|
| 44.1 kHz / 16-bit / stereo (CD) | 1 411 kbps | 10.1 MB | 605 MB |
| 48 kHz / 24-bit / stereo | 2 304 kbps | 16.5 MB | 988 MB |
| 96 kHz / 24-bit / stereo | 4 608 kbps | 33 MB | 1.98 GB |
| 48 kHz / 24-bit / mono | 1 152 kbps | 8.2 MB | 494 MB |
Recording a two-hour session in stereo at 96/24 costs about 4 GB. The same session at 48/24 costs 2 GB and, for anything other than heavy pitch manipulation, is indistinguishable in the final mix. Inspecting an existing file shows its actual sample rate, bit depth and channel layout — worth checking before assuming what a recorder or a download actually produced.
Mono Versus Stereo
A single microphone recording a single voice produces one signal. Storing it as stereo duplicates that identical signal into two channels and doubles the file for nothing. Recording interviews, podcasts and voice-overs in mono halves the size with zero loss — and if two identical channels are already in a file, folding them to mono is lossless. Converting the file handles both the channel layout and the sample rate in one pass.
Practical Settings
- Podcast or voice-over: record 48 kHz / 24-bit mono; publish as compressed audio, typically 96–128 kbps.
- Music production: record 48 kHz / 24-bit; use 96 kHz only if the project involves heavy pitch or time manipulation; deliver 44.1 kHz / 16-bit with dither.
- Video: 48 kHz throughout, 24-bit while editing, and whatever the delivery codec requires at export.
- Field recording with unpredictable levels: 32-bit float if the recorder supports it, otherwise 24-bit with peaks around −18 dBFS.
- Archiving: keep the original files at capture settings. Re-encoding an archive to save space destroys information you cannot get back.
Summary
Sample rate sets the highest frequency you can represent; 44.1 and 48 kHz already cover human hearing, and higher rates help during processing rather than during listening. Bit depth sets the dynamic range; 24-bit buys recording headroom, and 16-bit is entirely adequate for delivery. Record with room to spare, convert as few times as possible, dither once at the end, and match the destination rather than chasing bigger numbers.
