What are sample rate and bit depth?

Sample rate and bit depth in plain English

A microphone turns sound into a voltage that rises and falls with the air pressure. To store it digitally, the analogue-to-digital converter in your audio interface measures that voltage over and over and saves each measurement as a number. One measurement is a sample. The sample rate is how many samples it takes each second, and the bit depth is how many binary digits each of those numbers gets.

They control two separate things. Sample rate decides the highest frequency a recording can hold, and bit depth decides how quiet a sound can get before it disappears into noise. Neither one changes the loudest level a file can hold, which is 0 dBFS whatever the settings.

Playback runs the same process backwards: a digital-to-analogue converter turns the numbers back into a smooth voltage for your speakers or headphones. The lab at the top of the page is a converter you can set badly on purpose, from phone-call quality up to the 44.1 kHz, 16-bit format of a CD.

What is sample rate?

Sample rate is measured in hertz, meaning samples per second, so at 44.1 kHz the converter takes 44,100 samples every second, one every 22.7 microseconds. At 96 kHz it takes 96,000. In the lab's close-up screen every white dot is one sample, and the amber line through them is the sound that plays back.

On playback, the converter's output filter draws the one smooth wave that passes through every sample and contains nothing above half the sample rate. As long as the original had nothing up there either, what comes out is the wave that went in. That's the sampling theorem, worked out by Harry Nyquist and Claude Shannon, and it's why a higher sample rate doesn't make a 1 kHz tone any smoother.

The Nyquist frequency

Half the sample rate is the highest frequency a recording can hold, and it's called the Nyquist frequency: 22.05 kHz at a 44.1 kHz sample rate and 24 kHz at 48 kHz. A wave needs more than two samples per cycle to be pinned down, so anything faster can't be described by the samples at all.

People hear up to about 20 kHz when they're young, and the limit drops as they get older. A 44.1 kHz sample rate covers that whole range with about 2 kHz to spare, and the spare range is there for the converter's filter, which has to go from letting everything through at 20 kHz to blocking everything by 22.05 kHz.

Lower the rate in the lab and the dashed Nyquist line on the spectrum moves left. At 22.05 kHz it sits at 11.025 kHz, and the top of the drums loses its brightness while everything below stays as it was. At 8 kHz, the rate landline phone calls have used for decades, everything above 4 kHz goes and the loop sounds like it's coming down the line.

Aliasing and the anti-aliasing filter

Frequencies above Nyquist don't vanish on their own. At 44.1 kHz the samples of a 25 kHz tone are exactly the same as the samples of a 19.1 kHz tone, so the 25 kHz tone plays back at 19.1 kHz. That folding is called aliasing, and on a sound with lots of high harmonics, like a cymbal, it turns into a cloud of new tones that have nothing to do with the music.

To stop it, every converter removes everything above Nyquist before the samples are stored, with a steep low-pass filter called the anti-aliasing filter. The lab's filter passes everything up to about 250 Hz below the Nyquist frequency and is more than 80 dB down at Nyquist itself. Switch it off at 8 kHz and you hear what it protects you from: a metallic, out-of-tune layer, which is the sound a sample-rate reducer plug-in makes on purpose.

Aliasing can also happen inside your DAW with no converter involved. Distortion and clipping plug-ins create new harmonics, and the ones that land above Nyquist fold back down. The lesson on aliasing and oversampling shows how plug-ins deal with it.

Why 44.1 kHz and 48 kHz?

44.1 kHz is the sample rate of the CD, fixed in the Red Book standard that Sony and Philips published in 1980. The number comes from video. Before computer drives could hold an album, digital masters were recorded onto video tape through PCM adaptors, which stored three samples on each line of the picture. NTSC video has 245 usable lines per field and 60 fields a second, PAL has 294 lines and 50 fields, and both multiply out to 44,100 samples a second.

48 kHz came from professional and broadcast equipment, and it's the standard for anything with pictures: film, TV, online video and games. Its Nyquist frequency is 24 kHz, which gives the filter a little more room to work in.

For music, neither is audibly better, since both reach past 20 kHz. Use 48 kHz if the music is going to picture and whichever you like otherwise, then keep the whole project at that rate.

Do higher sample rates sound better?

For listening, no. A 96 kHz file holds frequencies up to 48 kHz, more than an octave above anything you can hear, and the part you can hear is the same as it is at 44.1 kHz. Recording at a higher rate can still help in a few situations:

• Distortion plug-ins alias less, because their harmonics have more room before they fold. Turning on the plug-in's own oversampling does the same job for less CPU.
• Latency is lower for the same buffer size: 256 samples is 5.8 ms at 44.1 kHz and 2.7 ms at 96 kHz, though the computer works twice as hard to keep up.
• A sound designer who records at 96 or 192 kHz can pitch a recording down an octave and still have content up to 20 kHz, as long as the microphone picked up something above hearing in the first place.

The costs are easy to work out. Every plug-in does twice the work at 96 kHz, and every file is twice the size: a minute of 24-bit stereo is about 17 MB at 48 kHz and 35 MB at 96 kHz. For most music, 44.1 or 48 kHz is the better use of your computer.

What is bit depth?

Each sample is stored as a whole number, and the bit depth is how many binary digits that number has. Every extra bit doubles the number of values a sample can take, so 16 bits allow 65,536 levels and 24 bits allow 16,777,216, spread evenly from the bottom of the wave to 0 dBFS at the top.

The real wave almost never lands exactly on one of those levels, so the converter rounds each sample to the nearest one. The difference between the true value and the stored one is called quantisation error. From one sample to the next it looks random, and you hear it as a faint, even hiss called the noise floor. More bits means smaller steps between the levels, and every bit you add pushes the noise down by about 6 dB.

That gives the rule of thumb for dynamic range, the gap between the loudest sound a format can hold and its noise floor: 6 dB per bit.

Lower the bit depth in the lab and watch the teal line on the noise floor screen climb. At 8 bits it's about −53 dBFS and easy to hear on the quieter moments. Drop to 4 bits on the sine wave and something else happens: with only 16 levels, the rounding error repeats with every cycle of the note and turns into distortion, a row of spikes at multiples of 110 Hz on the spectrum. That gritty, buzzy sound is exactly what a bitcrusher is for.

Is 24-bit better than 16-bit?

For listening to a finished track, 16-bit is already enough. It runs from the loudest sound a file can hold to a noise floor 96 dB below it, and a mastered track uses only a small slice of that range. A properly dithered 16-bit copy of a master sounds the same as the 24-bit file at any sensible listening level.

For recording, 24-bit is the right choice, and the reason is headroom. You can't know exactly how loud a singer or a drummer will get, so you leave space by recording with the peaks well under 0 dBFS, and every 6 dB of space leaves one bit unused at the top. Record peaking at −18 dBFS on a 16-bit system and you're using only about 13 of the 16 bits. At 24-bit there's so much room underneath that leaving plenty of headroom costs nothing you can hear.

The lab's last step shows the extreme. The guitar goes into the converter 48 dB too low and gets turned back up later, the way a too-quiet take is rescued in the mix. At 16 bits its peaks reach only about 8 of the bits and the hiss comes up with the guitar. At 24 bits the same mistake leaves the noise 48 dB lower, under anything you'd hear.

The 24-bit noise floor, about 144 dB down, is lower than any real equipment can reach. Microphones, preamps and the converter itself all add noise of their own well above it, so in practice the bottom few bits of a 24-bit recording are filled with the hiss of your gear and the room.

What about 32-bit float?

Inside a DAW, audio is processed as floating-point numbers, usually 32-bit and sometimes 64-bit. A 32-bit float has 24 bits of precision plus an exponent that slides the scale up and down, so it keeps the same fine detail at any level and has a range of well over 1,000 dB. That's why a channel can go over 0 dBFS inside the mix without clipping and only clips when it reaches a fixed-point output or file. The gain staging lesson covers what that means for your levels.

You can export 32-bit float WAV files too. They suit stems and bounces that are going back into a DAW, since a hot bounce can't clip, and some field recorders, including Zoom's F3 and F6 and Sound Devices' MixPre II range, record in 32-bit float so a take that comes in far louder or quieter than expected can still be used. For a finished master going to a distributor, 24-bit fixed point is the usual format.

Dither

Whenever the bit depth goes down, for example when a 24-bit master becomes a 16-bit file for a CD, every sample gets rounded again. On quiet material like a reverb tail or the end of a fade, that rounding stops behaving like random noise and starts following the music, which you hear as grainy distortion on the very end of the sound.

Dither fixes it by adding a very small amount of random noise before the rounding. The noise floor rises a little, about 5 dB with the common triangular kind the lab uses, but the error stops following the music and turns into plain hiss, which is much less noticeable. Add dither once, as the very last step, whenever the bit depth goes down. The dither lesson goes further, including noise shaping.

Reading the lab

Pick a sample rate with the buttons under the lab's title bar, and set the bit depth and the recording level with the knobs. The loops are 16-bit, 44.1 kHz files, so the lab starts at CD quality and can only go down from there.

The large screen is the spectrum, with low frequencies on the left and 22 kHz on the right. The pale shape is the original loop and the amber line is what comes out of the converter. When the sample rate is lower, the violet shading marks everything above Nyquist, which that rate can't hold, and with the filter off the pink fill is the aliasing. The teal line is the rounding noise. It's drawn one frequency at a time across thousands of frequencies, so it sits lower on this screen than its total level does on the noise floor screen.

The close-up shows a couple of milliseconds from a quieter part of the loop, about 24 dB under its loudest peak, where bit depth shows up first. Each dot is a sample. When the bit depth is low enough, faint horizontal lines appear: they're the levels a sample is allowed to land on, and you can see each dot snapped to one.

The noise floor screen shows the music as it arrives at the converter, from its quieter moments up to its loudest peak, with the noise floor as a teal line. Turn the recording level down and the music slides towards the line, while the line stays where it is. The lab measures the noise as an RMS level, which puts 16-bit at about −101 dBFS. The 96 dB figure comes from the 6 dB-per-bit rule of thumb; worked out exactly for a full-scale sine wave, the gap is 98 dB.

Switch on Hear only what's lost to listen to the difference between the original and the converted loop, at its real level. A/B switches to the original file, matched in loudness. You can also upload a file or record eight seconds of your own sound, and it stays on your device.

Which sample rate and bit depth should you use?

Neither setting changes how loud the track plays. Spotify, Apple Music and YouTube turn tracks up or down to a set loudness by default, measured in LUFS, and the LUFS lesson explains how that works.

How to set sample rate and bit depth

1. Choose the sample rate before you recordIn Ableton Live it's in the Audio tab of Live's settings, and in Logic Pro it's under File > Project Settings > Audio. Changing it halfway through means converting everything you've already recorded.
2. Check your audio interface agreesMost DAWs set the interface for you. If the interface's own control panel shows a different rate, change it there to match.
3. Record at 24-bitIt's the default in most DAWs. If you're not sure, look in your DAW's recording settings.
4. Leave headroom while recordingAim for peaks somewhere between −18 and −6 dBFS. At 24-bit that costs nothing, and it keeps a sudden loud note from clipping.
5. Export at the session's sample rateConvert only when the destination needs a different rate, such as 44.1 kHz for a CD from a 48 kHz session.
6. Dither only when the bit depth dropsSwitch it on for a 16-bit export, once, and leave it off for 24-bit and 32-bit float files.

Common mistakes

Mixing sample rates in one session

Most DAWs convert or resample files whose rate doesn't match the session. If one doesn't, a 44.1 kHz file played at 48 kHz runs 8.8% fast and about 1.5 semitones sharp, and a 48 kHz file in a 44.1 kHz session runs about 8% slow and the same 1.5 semitones flat. If a sample sounds slightly out of tune with everything else, check its sample rate first.

Upsampling to improve a file

Converting a 44.1 kHz file to 96 kHz adds nothing. The frequencies above 22.05 kHz were never recorded, so the new file is just twice as big, and turning a 16-bit file into 24-bit only adds empty bits at the bottom.

Recording too low at 16-bit

If your interface or recorder is set to 16-bit, every 6 dB of unused headroom costs a bit, and a take recorded far too low brings its hiss up when you turn it up in the mix. Switch to 24-bit and the same mistake becomes harmless.

Confusing bit depth with bit rate

Bit rate, in kilobits per second, is how much data a file uses every second. A 320 kbps MP3 has a bit rate, but it doesn't have a bit depth in the same sense, and choosing 24-bit in an MP3 export window doesn't make it a 24-bit file.

Dithering twice

Mastering limiters often have their own dither switch, and so does your DAW's export window. Using both adds two layers of noise, so pick one and turn the other off.

Training your ears to hear sample rate and bit depth

The quiz under the lab plays two versions of the loop at the same loudness, with the screens covered. One is the original file and the other has been converted to a lower sample rate or bit depth, and you pick the converted one. There are eight rounds, and your best score is saved in your browser.

Questions people ask

Keep going