Why 24 Bit Matters and 96 kHz Does Not

This page contains affiliate links. If you buy through one we may earn a commission, at no cost to you.

Almost never, and the reason is that the question treats sample rate and bit depth as two dials on the same quantity. They are not. One of them has a hard ceiling that the ear sets, and 48 kHz is already above it. The other decides how wrong you are allowed to be about the gain knob, and that is a real thing to buy. So the common default, 24 bit at 48 kHz, is the right one, and the common mistake is the exact reverse: 16 bit at 96 kHz, which doubles your file to capture nothing and keeps the one constraint that can actually spoil a take.

Nyquist, and why 48 kHz is already finished

Sampling theory gives you one clean result. A stream of samples taken at a rate of R captures, completely and without loss, every component of the original signal below R divided by two, and nothing at all above it. That half rate figure is the Nyquist frequency, and it is not an approximation or a rule of thumb. Below it the reconstruction is exact. Above it there is no information at all, and worse than none, because any content that was up there folds back down into the audible band as a false low frequency, which is why every converter has a filter in front of it whose whole job is to remove the material it cannot represent.

So 48 kHz reaches 24 kHz. A healthy young ear reaches about 20 kHz and most adult ears stop between 15 and 16 kHz. Moving to 96 kHz raises the ceiling to 48 kHz, which means the entire purchase is the span from 24 kHz to 48 kHz. Nothing in a human voice lives there. No microphone diaphragm we carry is specified to produce it. No ear at any age receives it. You are paying to record an octave of silence above an octave nobody can hear.

Put the cost in terms of a card we sell with a recorder. A 32 gigabyte card holding 24 bit stereo at 48 kHz carries 288,000 bytes per second, so 32 billion bytes divided by 288,000 is about 111,000 seconds, which is 30.9 hours. Switch to 96 kHz and the rate doubles to 576,000 bytes per second, so the same card holds 15.4 hours. You have halved your recording time, and in return the file now contains a frequency band that is empty by physics.

There are two honest exceptions and both are small. If you intend to pitch a recording down by a large factor, the ultrasonic content becomes audible content after the shift, which is why sound designers record wildlife and machinery at high rates. And a higher rate lets the anti aliasing filter be gentler, because it has further to roll off before it matters, which removes a little phase distortion near the top of the audible band. Neither of those is a reason for the recordings these products are bought for, which are voices at conversational level.

Bit depth is a different quantity, and it is the one that matters

A sample rate sets the frequency ceiling. A bit depth sets the distance between the loudest thing you can represent and the quantisation noise underneath everything. Each bit doubles the number of levels, and a doubling in amplitude is 20 times the base ten logarithm of 2, which is 6.02 decibels. So 16 bit gives you 16 times 6.02, which is 96.3 decibels, and 24 bit gives 144.5.

Those numbers only mean something next to a microphone, and here our own catalogue supplies both ends of the window. The Rode record publishes a self noise figure, and the Zoom record publishes a maximum level. Reference the system so that the Zoom's published 120 decibel maximum lands at digital full scale, and work out where each floor actually sits as a sound pressure level in the room.

At 16 bit the quantisation floor is 96.3 decibels below full scale, so it sits at 120 minus 96.3, which is 23.7 decibels of sound pressure level. The microphone's own noise is at 15. So the converter is the louder of the two noise sources, by 8.7 decibels, and 16 bit is genuinely the limiting component.

At 24 bit the floor is 144.5 decibels down, so it sits at 120 minus 144.5, which is minus 24.5 decibels of sound pressure level. That is not a quiet noise. It is a level below the thermal agitation of air molecules, which means no microphone can ever reach it and no recording will ever contain it. The converter noise has stopped existing as a consideration, with 39.5 decibels to spare.

Which locates the whole decision precisely. The useful crossing point lies between 16 and 24 bit, the microphone is on the wrong side of 16, and there is no third option on the market. That is the entire argument for 24 bit and it is sufficient on its own.

What the extra eight bits are really for, which is permission to be careless

Everything above assumed you set the level perfectly, with the loudest moment of the take landing exactly at full scale. Nobody does that, because the penalty for overshooting is a clipped and unusable recording while the penalty for undershooting looks like nothing. So the standard practice is to aim low, commonly with peaks around 18 decibels below full scale, and lift the whole thing afterwards.

Aiming 18 decibels low moves both digital floors up by 18, because you have simply used less of the available range. The 16 bit floor goes from 23.7 to 41.7 decibels of sound pressure level. A quiet domestic room measures around 30. So the converter noise is now 12 decibels louder than the room you were trying to record, and when you lift the file by 18 decibels in the edit the hiss comes up with the voice and arrives plainly audible. The 24 bit floor goes from minus 24.5 to minus 6.5, which is still unreachable, and the lift brings up nothing.

That is the real purchase. Not detail, not warmth, not resolution. Eighteen decibels of permitted error on a knob you have to set in a hurry, before a take you cannot repeat. Twenty four bit at 48 kHz is the sensible default because it spends its extra bits on the one thing that actually goes wrong.

What 32 bit float is, and what it is not

Two of our recorders advertise a bigger number still, and it is worth being precise about it, because the number is not doing what the name suggests.

A Zoom H1essential handy recorder bundle leads on 32 bit float, and our own record of it repeats the claim as published: 32 bit float recording eliminates level adjustments, with built in X and Y microphones capturing up to 120 decibels of sound pressure level, plus a 32 gigabyte card, a lavalier microphone, a windscreen and batteries. The 120 decibel figure is the one worth having, because it is half of the window the drawing above needs and nothing else we carry publishes it.

Now the format. A 32 bit IEEE float is not 32 bits of resolution. It is a 24 bit mantissa with an 8 bit exponent, so its precision at any one scale is 24 bits, which is the same 144.5 decibels as a 24 bit integer file. What the exponent adds is the freedom to slide that 144 decibel window anywhere across an enormous span, which means the file cannot clip and cannot run out of room at the bottom. So the format contributes no extra resolution whatsoever. It contributes a container that does not care where the level landed.

The uncliippability therefore cannot come from the file format, and it does not. The technique these recorders use is to run two analog to digital converters at different gains at the same time and combine their outputs, so that the quiet converter handles the soft passages and the loud one survives the peaks, and the result needs a format with enough exponent range to hold the join. Neither our Zoom record nor the record of the other 32 bit float product we carry says any of that, which is the gap worth reporting: both listings attribute to a number what is actually done by a pair of converters. No converter on the market has 32 real bits, and the best measure out at roughly 21 effective bits, because at that point you are reading the thermal noise of the resistors in the input stage.

None of which makes the feature fake. It is genuinely useful and it genuinely means you can stop watching a meter. It just is not resolution, and if you were hoping the number meant your recordings contain more of the world than a 24 bit file would, it does not.

Which of our listings publishes enough to answer the question

Three products publish the right pair of figures and the rest publish neither, which is worth laying out plainly.

A Hollyland Lark M2 wireless microphone publishes 48 kHz at 24 bit, and that is exactly the default this page argues for. Its record puts the figure under a heading about studio grade sound, which is marketing, but the figure underneath it is the correct engineering choice and deserves to be said out loud: the sensible answer was already on the box. The rest of the record is about the things that genuinely vary, with 9 gram transmitters, 10 hours per charge and 40 with the case, a 300 metre range and one button noise cancellation, and it is honest that the noise cancellation reduces background noise rather than removing it and that battery life drops when it is on. Full record in our Hollyland Lark M2 review.

A BOYA mini 2 wireless lavalier microphone publishes the same 48 kHz at 24 bit, along with a 5 gram transmitter, a limiter and a safety track. A safety track is worth a sentence here because it is the same idea as 32 bit float solved the older way: record a second copy at a lower gain, so that if the main one clips you still have a usable take. Two of the four products on this page are carrying a defence against level error, which tells you what the real failure mode is. See our BOYA mini 2 review.

And the one listing on the site that publishes the honest limit. A Rode VideoMic GO II-H shotgun microphone bundle publishes class compliant 24 bit at 48 kHz over USB, a supercardioid pattern, dual 3.5 millimetre and USB-C outputs and no battery, and then it publishes the figure that decides everything above: a self noise of 15 decibels A weighted. Our own record treats that as a drawback, listing it as a self noise of 15 dBA that is not the lowest and recording as its main weakness that 15 dBA may be audible in quiet settings. Both readings are fair. The point this page adds is that the same figure is also the most useful number in the category, because without it you cannot tell whether a bit depth is doing anything. One product in nine publishes it. Details in our Rode VideoMic GO II-H review, and the rest of the category is in our microphones for cameras guide.

A sentence of ours that gets this exactly backwards

We should correct our own page before anyone else's. Our Bluetooth receivers guide says of one pick that "a built-in 384kHz/32-bit upsampling DAC aims to preserve detail that simpler adapters gloss over", and lists a dedicated audiophile grade converter among its advantages.

Take the sampling theorem at its word and that cannot be true. The receiver's input is a Bluetooth stream, and the codecs the same record names carry 48 kHz for aptX HD and up to 96 kHz for LDAC. Whatever arrives has already been sampled, and by Nyquist it contains nothing above half of its own rate: for a 48 kHz stream, the content above 24 kHz does not exist to be preserved. Upsampling to 384 kHz then computes seven interpolated values between every pair of real ones, and every one of those values is already fully determined by the samples on either side. The information added is zero bits. Not a small amount. None.

Upsampling is a real and worthwhile technique, and the reason is unglamorous: it moves the reconstruction filter's work to a frequency where a simple analog filter can finish the job, which reduces distortion in the converter's output stage. That is a benefit in the converter, not a recovery of detail from the stream. The listing's own hedge, "aims to", is doing more work than it looks like. And the honest reading of the same spec sheet is the one the record already gives elsewhere, that this is a receiver with no amplification, which is the sort of limitation worth checking before any of the numbers matter.

Problems in our own records

Two things turned up that are worth reporting rather than burying.

The first is a gap. Of the nine microphone and recorder products we publish, one gives a self noise figure, three give a bit depth, three give a sample rate, and one gives a maximum sound pressure level. Not one gives both a self noise and a maximum level, so no single listing on this site contains enough to state its own dynamic range. The two figures the drawing above needs had to be taken from two different products made by two different companies. That is a category wide absence rather than anyone's failing, and it is why a buyer cannot answer this question from a spec sheet.

The second is a defect in our own database and it is larger than previously recorded. The specification fields of all four products named on this page are corrupted in the same way: labels reading only "Feature", or truncated in the middle of a word, with values cut off mid sentence, and the advantages stored as one unbroken run of text. We had logged this for four dishwashers and one electric kettle. Running a detector for the signature across the whole catalogue finds it in 286 of the 401 published products whose identifiers fall between 3500 and 3900, and in none at all outside that range. So it is not a handful of bad records. It is one import cohort, about seven products in ten inside it, spanning cameras, lenses, lighting, air fryers, espresso machines, telescopes and the microphones here. Every figure quoted on this page was read from the summary and the advantage and drawback lists instead, which are intact.

The honest limit, and the free technique that beats the purchase

The specification that would settle this is the microphone's own noise floor and its maximum level, as a pair, and the category does not publish them. So the realistic position is not that you should hunt for a better number. It is that the numbers on offer describe the converter, the converter stopped being the limit at 24 bit, and the component that is the limit is silent about itself.

What follows from that is a short and slightly dull set of defaults. Record 24 bit. Record 48 kHz unless you have a specific reason involving pitch shifting. If your recorder offers 32 bit float, use it and stop watching the meter, because that is a genuine convenience. If it does not, aim your peaks about 18 decibels low and lift afterwards, which is what the 24 bit was for.

And then the move that costs nothing and beats every number on this page. Sound pressure falls with distance, so halving the distance from a source raises the level by about 6 decibels, which is one whole bit of dynamic range. Move a microphone from 24 inches to 6 inches and you have halved the distance twice, gained 12 decibels, and bought two bits for free. The entire step from 16 bit to 24 bit is eight bits, most of which no microphone can use. Eighteen inches of repositioning delivers two bits that every microphone can use, because it raises the thing you want without raising the room noise, the air conditioning or the traffic outside, none of which moved closer.

It compounds too. The same change improves the ratio of direct sound to room reflections, which is the single biggest difference between a recording that sounds professional and one that sounds like a room, and no amount of bit depth touches it. One caution, and it is the subject of a companion page: on a directional microphone, moving in this close changes the tone as well as the level, by a large and predictable amount. That is the proximity effect, and it is worth understanding before you take this advice literally. If you want the wider argument about why a bigger number on a specification rarely survives contact with physics, our piece on why 100 watts is not twice as loud as 50 makes the same case in the acoustic domain.