Does the resolution of an audio recorder really matter?

[Emre Safak]

After buying a 24/96-capable Marantz recorder I just had a thought. Does the resolution on these things even matter when the SNR is limited by the pre-amp's and mic's internal noise? How low does the resolution need to be before it becomes the bottleneck, say with a good mic in the field?

[Wayne Brissette]

Emre:

I think ultimately it depends on what type of work you are doing. I do everything from docs to orchestras, and do it all with a single recorder. Prior to the Zaxcom Deva, I carried around a mixer, DAT recorder, external A/D converters, all sorts of things. Basically it was a mess. For voice/docs I could get away with the 16-bit limitation of the DAT recorder, but for music, I used an external A/D which captured a 24-bit signal and dithered it, so it ended up as a 16-bit signal before going to the DAT recorder. Some of those recordings have stood up quite well to the all 24-bit recordings I'm doing today.

If you A/B a pure 16-bit recording with a 24-bit recording done at the same time, under identical conditions, you'll be able to hear a definite difference. Where things get fuzzy is the sampling rate. Going from 44.1 kHz to 48 kHz or 96 kHz doesn't always buy you a better recording. In fact several years ago, I took part in a blind test, where several people who claimed they could tell the difference were giving the identical recording with 16 and 24-bit recordings done at the same time. That change was pretty easy to detect, but few (I don't think anybody actually) could consistently pick out from 24-bit recordings when the sampling rate changed. So, certainly there is some validity to the SNR getting better on new equipment and making it less of a point to record at a higher sampling rate. That said, I really would enjoy doing the test again with Sanken's 100 kHz microphone. I don't know if it really would make any difference, but it would be a nice test to take part in anyhow.

Wayne

[Emre Safak]

That's another thing. I totally forgot that most mics top out at 20khz... making high sampling rates seemingly all the more ridiculous. Is it more about giving yourself more leeway in the DAW?

[Mike Peter Reed]

I've yet to do my unscientific testing on my recently acquired 702T. Previously I used a Fostex FR2 and could swear there was a difference between recording (of all things) dialogue at 48K and 96K (both at 24-bit). Really really marginal, but somehow the 96K recordings sounded more "real" on playback. Now I've never recorded 192K except for sound effects that might get stretched around, but 96K became my standard just based on the FR2 experience, downsampling to 48K if production requested. Must try the same again on the 702T because I've also read somewhere that certain A/D convertors work better at higher sampling rates. So, if the 702T has far superior A/D than the FR2 then recording at 48K may sound as good as recording 96K on the FR2. If you see what I mean ....

[Steve House]

Originally Posted by Emre Safak

That's another thing. I totally forgot that most mics top out at 20khz... making high sampling rates seemingly all the more ridiculous. Is it more about giving yourself more leeway in the DAW?

If you are mixing digitally the noise floor increases as additional sources are added into the mix. This increased noise floor is the equivalent of a recording at lower bit depth. Losing the equivalent of (just for example) 4 bits due to this effect means that when mixing several 16-bit sources you end up with the equivalent of a 12-bit S/N ratio while mixing the same number of 24-bit sources still leaves you with the equivalent of a 20-bit recording. While the difference between a 16-bit recording and a 24-bit recording may not be audible, the difference between a 12-bit and 20-bit final mix is dramatic. Holman discusses this effect more in-depth in his text "Sound for Film and Video."

[Emre Safak]

... but doesn't the DAW process the audio internally at a higher resolution, like 32-bit float? As long as you acquire with a high enough resolution to exceed your equipment's SNR, shouldn't you be fine?

[Marco Leavitt]

There seems to be consensus among sound people that 24 bit recording is worth it if your recorder is high enough quality to really take advantage of it (most aren't), but 48K is plenty for dialog, and 96K isn't worth the extra storage space. I would think you would want to use a higher sampling rate for orchestras and things like that though.

[Martin Pauly]

Originally Posted by Emre Safak

I totally forgot that most mics top out at 20khz... making high sampling rates seemingly all the more ridiculous.

I have heard/read this argument before, and while my ears agree based on the recording work that I have done, I still wonder if there isn't an advantage in describing the waveform of, say, a 20kHz note with more detail/precision?

Taking it to an extreme, assume a 24kHz tone recorded with 48kHz sampling rate, how could one distinguish between a pure sine wave and other possible shapes of this note? Is the difference not important because our ears can't tell them apart?

I am not an expert in this area; this is just something that I've been thinking about.

- Martin

[Mike Teutsch]

Originally Posted by Emre Safak

That's another thing. I totally forgot that most mics top out at 20khz... making high sampling rates seemingly all the more ridiculous. Is it more about giving yourself more leeway in the DAW?

I'm no expert, but I think you are confusing the mics maximum frequency response with the recorder sampling rate. Very different animals.

The mic outputs in linear and the digital recorder sample in different rates, as set up by you. The more often it is sampled, as 12 vs 96, the more natural the sound or the more information you have to work with.

Mike

[Emre Safak]

Originally Posted by Mike Teutsch

I'm no expert, but I think you are confusing the mics maximum frequency response with the recorder sampling rate. Very different animals.

Not at all; they are closely related through sampling theory. If your mic's frequency response tops out at 20khz, sampling the sound at 44.1khz or 48khz should be enough to capture all the information.

[A. J. deLange]

With respect to bit depth it works this way: if you have n bits to play with one is taken up specifying whether the sampled voltage is positive or negative. Two more are taken up by headroom i.e. an A/D presented a complex (relative to a tone) input will start to overload when the rms input voltage is around 12 dB (approximately - the actual value depends somewhat on the number of bits) below the rail (the rail is the highest DC voltage which the A/D can represent). The quantizing noise is about 2 bits (10.8 dB) below the least significant bit's power. If you set the noise floor of the source (mic/preamp combination) at the level of the LSB the quantizing noise is then insignificant relative to the source system self noise (the desired condition). You wouldn't want to record within 10 or so dB of the mic/preamp's noise floor so you set things up (mic type, location, preamp gain...) so that the softest sound you record is say 12 dB (2 bits) above the LSB. Thus any level that is more than 2 bits above the LSB and more than 2 bits below the rail is at least 10 db above the input system (mic/preamp) self noise and below the distortion level. You have, therefore, n - 1 - 2 -2 = n-5 bits of dynamic range which is 6*(n-5) dB and, if you know the dynamic range of the sounds to be recorded you can figure out the required n. If 6*(n-5) is greater than the dynamic range of the source adding extra bits doesn't do a thing for you. This is, apparently, a hard concept to grasp but nevertheless true. Another thing to consider is that a 24 bit A/D converter is most unlikely to have 6*(24 - 16) dB more dynamic range than a 16 bit A/D converter. We speak, in such cases, of "effective bits". The number of effective bits is the dynamic range divided by 6 plus 5 (with dynamic range defined the way I've defined it here which is somewhat arbitrary). In any case if A/D converter 1 has 9 dB better dynamic range than A/D converter 2 it has 1.5 more effective bits.

When it comes to processing consider multiplying 10.5 by 7.1 = 74.55. If we are only allowed 1 decimal place we must truncate to 74.5 or round to 74.6. In either case the error is less than the least significant digit. It is much the same with binary arithmetic. The number of bits lost is more like one or 2 than 4 or 5 if things are done with care. The proper approach is to carry as many bits as you can until the last minute and then round. Floating point does help with this.

When it comes to rate, intuitive as it may seem that more samples per cycle would convey more information than 2 this is not the case. Two samples of the highest frequency to be conveyed are sufficient. This is a theoretically correct statement but there are practical considerations. The input signal must be "strictly band limited" which is also a theoretical condition that cannot be met in practice but which can be approximated sufficiently closely. The problem lies in the analog filter which must precede the A/D converter. It is very hard to build an analog filter which doesn't attenuate up to 20 KHz, has flat delay and amplitude response up to that frequency and yet rejects signals strongly at 22.05 KHz (half 44.1 KHz) and above. It's a little easier to do this for a sampling rate of 48 KHz (needs to attenuate strongly above 24 KHz) and easier still with 96 and 192 MHz rates. What is easy to do is come up with a relatively simple filter which cuts (starts to roll off) at 20 KHz and is way down at 96 KHz), A/D convert and then use a linear phase (FIR) filter to low pass filter to 20 KHz with stop band starting at 22.05, 24, or 48 KHz. The outputs of these digital filters can then be decimated without introduction of aliasing. If I had to build an A/D box capable of sampling at 192 kHz this is exactly how I would do the lower sampling rates and I'm guessing that I'm not the only guy in the world that knows a little DSP. If I'm right that the manufacturers do it that way then there shouldn't be a detectable difference in the quality of the sound at the higher sample rates (given, of course, that the reconstruction process to which similar considerations apply, does an equally good job for any sampling rate). The proof of the pudding WRT to any particular system is in a double blind triangle test.

[Petri Kaipiainen]

I try to state my view as simple as possible:

24 verus 16 bits: there is a true benefit in using 24 bits for recording, even though 16 bits is more than needed in real life (hardly anybody has systems and listening rooms to use up even 16 bits/96 dB of dynamic range). This is because it gives more safety and headroom in recording and mixing. An example: I took the PA signal from a soundboard in a hotel ballroom setting the levels from the signal given by the sound person (real signal, not tone). I had no possibility to readjust during the show. To be safe I adjusted the signal to lowesh level on my SD722 using 24 bits. The end result was low in level, max about -25 FSdB. It was perfectly safe to raise levels in post without any noise, the end result was just as good as 16 bits with perfect levels, as the SD722 has s/n ratio of around 110 dB.

Using higher sampling rates is not usfull at all. Waveform "detail" which somebody here mentioned in this thread is nothing more than higher frequences. Because we can not hear them, mics can not pick them up and speakers can not reproduce them they are not needed. It is funny some people advocate 24/96 recording when these goals are impossible to achive at the same time: making a mic which hears 40-50 kHz range means the membrane is small. Small membrane mic has bad s/n ratio, less than 70 dB, which is only 12 bits... So it is one or another, never both at the same time. 96k sampling is usefull only with effects, AND only with very special microphones.

Of course using 96kHz sampling does no harm, just uses up disk double the space, but nobody should belive he gets more audible quality out of that.

[Emre Safak]

Thanks for confirming my suspicions. I'll record at 24/48 from now on and reserve higher sampling rates for mastering.

[Mike Teutsch]

Lots of very interesting info here. Thanks for the education. I'm reading and rereading.

Mike

[Seth Bloombaum]

I'd like to reinforce what Petri wrote - his opinions and practices are shared by me and several pros I know and work with.

For certain, 24-bit is a lifesaver when quieter program comes in, expected or unexpected. It is a great reassurance that when you hear something you want, and realize that it went in at -32db, there will be plenty of signal and no audible digital noise floor because you're recording in 24-bit.

Same on 48KHz - higher sampling rate doesn't get you much if anything.