Bit depth

Bit depth is dynamic range, not resolution#

The number of bits is the information you have available to represent the dynamic range, meaning the distance between the loudest thing in your music and the quietest thing, or the signal and the noise floor. If you only take away one thing from this page, it should be this: the quieter the sounds are, the more bits you need to capture them in higher quality.

This is where most people go wrong. CDs came out in the 80s, digital video from the 80s looks awful, 4K has more pixels and looks better, so it must be the same with audio: more bits, more resolution, better sound. No. Bit depth is not pixels. Quiet signals suffer disproportionately at low bit depths, and loud signals are privileged. More bits do not make loud material sound better.

On loud music, 8 bits can be enough#

Contrary to popular belief, just 8 bits might sound perfectly fine if you are listening to loud, dense, noisy music. You might not even be able to tell the difference between 8 bits and 32 or 64. The second you turn the music down and listen to something quiet with a lot of dynamic range, say an orchestral recording, you immediately hear that 8-bit isn't going to cut it.

So it seems obvious where this goes: if 16 is better than 8, then 24 must be better than 16, and 32 better than 24, and 64 better still. Well, no.

Your ears and your converter set the ceiling#

The first limit is the human ear. It has roughly 120 dB of usable dynamic range, helped by a kind of mechanical compression in the bones and muscles of the ear, and about 140 dB is just pain.

The second limit is the converter. Manufacturers mostly buy their chips from the same big four semiconductor makers, and those chips are rated around 110 to 120 dB. The best I had found at the time of making my video was 132 dB, which is less than 22 bits (see DACs). But interestingly in a similar way to what Douglas Self points out in his book Small Signal Audio Design discussing opamps, it is possible to run multiple channels of conversion on one chip in parallel, for the same mono signal, and the signal to noise ratio can improve further than each chip individually. ESS's flagship DAC is rated at 140 dB in mono mode, with all eight of its channels paralleled, against 132 dB per channel. That chip was announced in 2016, so it already existed when I made the video; I just hadn't come across it. But it doesn't change the conclusion that even at an insane 140 dB, that's still less than 24 bit dynamic range.

16-bit, correctly dithered, goes further than people think. It can reproduce low-distortion sine waves more than 100 dB below full scale, below the noise floor of the dither itself. At a reasonable listening level, on quality equipment, 16-bit already borders on the limits of human hearing.

24-bit fixed point is already overkill#

That is why 24-bit fixed point is still the studio gold standard. It is already more than you need for recording and for delivery, and the extra bits of the bigger formats sit below what your converter can produce and what your ears can hear. You don't need more than that.

32-bit float is 24-bit with a volume knob#

So what is floating point? Imagine you could take your resolution and slide it up and down like a volume knob, so quiet passages get the same precision as loud ones. That is floating point. In 32-bit float, eight bits are the exponent that does the sliding, and the actual numbers part is the same 24 bits as regular 24-bit.

Unless you are making music so quiet that nobody can hear it, float's huge range buys you nothing.

Float is important inside the DAW#

If you take something really quiet, turn it up really loud, turn it back down again and keep doing that through a chain of plugins and wild gain staging, float prevents the accumulation of quantisation error. That is why a lot of DAWs run 64-bit float internally, and this is good.

Float also lets you go over 0 dBFS, way into the red, then turn it down afterwards and get the original back with no distortion. It is fun the first time you do it, but it is a bit of a gimmick; you don't really need to do it. Clipping a fixed-point file is a different story, because that data is then gone; see clipping.

Better mixing beats more bits#

Any quality gain you think you will get by upgrading your bit depth is absolutely dwarfed by the quality gain you will get from upgrading your mixing skills.