44khz,48khz, 96khz, 192khz Spectral Frequency Comparison

[DavidNJ]

These are spectral frequency displays .5 second of me speaking that was recorded at 1-bit 2.8Mhz and then exported to each of the these settings in order:

44khz@16 bit
48khz@16 bit
48khz@24 bit
96khz@24 bit
192khz@16 bit
192khz@24 bit
192khz@32 bit floating

I clipped the pictures of 96khz and 192khz at 28khz. Noise began at around 32khz.

What I found unexpected the increase in frequency sharpened the view in the audible ranges while the difference in resolution didn't. Can someone explain that?

 

[Noiz2]

Interesting but how did you record them? Seven identical recorders fed from a distribution amp and a single microphone? That's about the only way you could get results with meaningful information. It's still interesting but if there wasn't a rather extreme control of the variables it's hard to tell if were looking at converter differences or converter at different sample rate differences or maybe the way whatever you were creating the spectrograph with images differently at different sample rates??? Added definition at lower frequencies has not been one of the claims of the pro high sample rate crowd so apparently nobody's hearing this.

 

[DavidNJ]

There was capture to 2.8Mhz pulse density modulation on a Korg MR-1 and used Korg's AudioGate tool to render the same master at the different resolutions.

The hypothesis was that 16 vs. 24. vs 32-bit encoding would show a difference, not the sampling frequency. The spectral analysis seems to show just the opposite; difference based on sampling frequency and minimal visible based on sample resolution.

 

[JordanBlock]

There was capture to 2.8Mhz pulse density modulation on a Korg MR-1 and used Korg's AudioGate tool to render the same master at the different resolutions.

The hypothesis was that 16 vs. 24. vs 32-bit encoding would show a difference, not the sampling frequency. The spectral analysis seems to show just the opposite; difference based on sampling frequency and minimal visible based on sample resolution.

It seems to me that all you can really read from this is the quality of the down sampling algorithm you used at various bit depths and sample rates.

 

[DavidNJ]

Maybe...that is easily tested...I can down sample the 192/32 to 48/24 or 44/16 with a different tool...and I will post tomorrow, but no, it looks the same.

This may be a filter issue. The frequencies are well below the Nyquist limit yet they seem to be more sharply resolved temporarily by the higher sampling frequency. It may be a low pass filter truncating information below the limit in an effort to limit higher frequency artifacts.

 

[Noiz2]

Yes this doesn't say much about sampling since your converting not sampling. You might try playing the audio analog out of the MR-1 and into some respectable A/D and recording that at different sample rates.

OT how do you like the MR-1? I've heard good things but haven't actually touched one yet. I'm not a big advocate of high sample rates but if your going to go there them the MR-1 (or some other 1bit recorder) seems to me to be the way to go.

 

[DavidNJ]

Once it is a digital PDM there is no reason to ever go analog. Korg provides software to deliver PCM, which is how I generated the first conversions. Then took the 192/32 and converted it in Audition, just as you would for output to CD/DVD.

I am blown away by the MR-1. It is end of life and has no XLR, but am considering getting a second. It just sounds clearer. I've found it necessary to delete the signal above 30khz...but below it the sound is different than what I am used to.

The difference I think is an MTF similar to what we see in pictures. As you get close to the Nyquist, a low pass filter set to reduce reflection reduces MTF. I am pretty sure that is what is happening in the 44/48khz signals which have Nyquist limits barely out of the audiable range, and witih in the range (with distorition) of the microphone. 96khz raises the Nyquist probably enough to fix most of the problem. However, it seems 192 raises the bar higher.

Basically, all that noise above 30khz isn't lost. It is reflected back and misrecorded below 30khz. Capturing and deleting it makes a cleaner signal.

I've tried to post a few times with graphs from Canon and Panavison to show the same thing with optical images. The reason the Canon 5DM2 looks great is that it oversamples. The MTF stays high right though 400lp/ph horizontal. The EX1 is lossing by 200lp/ph and is just above noise at 400.

Right now, I'm a convert. 96kHz is the lowest I will record, 192 is better, 1-bit best, and I don't care about the bit depth.

 

[Noiz2]

Once it is a digital PDM there is no reason to ever go analog. Korg provides software to deliver PCM, which is how I generated the first conversions. Then took the 192/32 and converted it in Audition, just as you would for output to CD/DVD.

That depends on what your trying to find out. If your looking for a reason to record at higher SR then you have to take a signal through the path it will go to get recorded. In other words if your trying to convince me that the mids will be cleaner at 96 then you need to record through the A/D converters to a recorder at 96. As it stands now you nice pictures of a work flow that VERY few people use. If your trying to show that you get tighter mids by converting MR-1 recordings to higher sample rates well fine but that's not what most people are recording with. It's very hard to set up an experiment that actually has meaningful results and very easy to create one that supports what you want to believe.

Basically, all that noise above 30khz isn't lost. It is reflected back and misrecorded below 30khz. Capturing and deleting it makes a cleaner signal.

Reflected back from what? Is this based on any real tests or just what you think is happening. It would mean that the brick wall filter was rather poorly designed since the hifreq. get filtered out before the A/D converter. Or are you talking about your down sampling software?

Right now, I'm a convert. 96kHz is the lowest I will record, 192 is better, 1-bit best, and I don't care about the bit depth.

Hmmm. In a whole lot of blind listening tests we did everyone could spot 24 VS 16 bit recordings, hardly anybody could hear a difference between 96 and 48. Many people seem to think that recorders capable of higher SR sound better even at 48k. This makes a certain amount of sense since the converters are generally of better quality and these machines may be doing an internal down sample to get the lower SR. In effect doing sort of what your doing.

If your a convert based on graphs I'm sorry but I can't put too much credence on that. And this is a list for filmmakers not insect mating call recordists. I mean you can record at what ever you want but if your going into a film work flow the first thin audio post is going to do (after cursing the stupid mixer that recorded everything at 96) is bulk SR convert your tracks down to something usable like 48. If your recording classical music or some sound FX then yes but production...

 

[j1clark@ucsd.edu]

Reflected back from what? Is this based on any real tests or just what you think is happening. It would mean that the brick wall filter was rather poorly designed since the hifreq. get filtered out before the A/D converter. Or are you talking about your down sampling software?

I believe the technical term for that is 'frequency aliasing', and is why one puts a 'sharp', or as 'sharp as possible low pass filter with cut off at frequency X for a given sampling rate Y (like 1/2 the sampling rate...), so as eliminate or mitigate such aliasing.

So, for 44 KHz sample rate, that would be a low pass filter with 22 KHz as the cut off...

 

[Noiz2]

I believe the technical term for that is 'frequency aliasing', and is why one puts a 'sharp', or as 'sharp as possible low pass filter with cut off at frequency X for a given sampling rate Y (like 1/2 the sampling rate...), so as eliminate or mitigate such aliasing.

So, for 44 KHz sample rate, that would be a low pass filter with 22 KHz as the cut off...

Yes, but where is the reflecting happening? And I think you have your chain wrong. The 1/2 freq is because it takes at least two points on a curve to extrapolate a freq. The brick wall filters that your talking about would be at about the sample rate and before the converters. They are there to eliminate higher feq. that would confuse the converter (ie anything that wouldn't have at least two samples on the same wave). The filter that is about 1/2 the SR is after the converter and is there to smooth out the converted wave and remove the stair steppienes. One legit reason why higher sample rates "should" sound better is that even if you can't hear above 20k (more like 10-15 for most adults) the feq at the top are now better described by having more data points on the wave. My experience is that pretty much nobody does hear this extra detail unless your recording VERY bright sounds and even then it's pretty subtle.
My problem with recording everything at 96+, other than compatibility, is a bang for buck issue. It is a significant hassle working at 96+ and in most circumstances brings no perceptible benefit.

 

[DavidNJ]

The higher frequencies are immutable after the A/D conversion. you need to remove them before. The photo sites use an optical filter.

This may be a contrast issue. Even if the frequencies being observed is less than 24khz, the sampling of the amplitude at the frequency could be more frequently. Nothing forces multiple sounds to be phase aligned. Just a thought.

 

[Noiz2]

The higher frequencies are immutable after the A/D conversion. you need to remove them before. The photo sites use an optical filter.

This may be a contrast issue. Even if the frequencies being observed is less than 24khz, the sampling of the amplitude at the frequency could be more frequently. Nothing forces multiple sounds to be phase aligned. Just a thought.

Not sure what your talking about. Higher frequencies have to be removed prior the the A/D because they will freak out the converter. With only one point on the curve the converters guess at the feq. will be all over the map. No idea what you mean by "immutable". That would imply that you could not filter after A/D which is of course not the case.

And yes all frequencies ARE sampled more with higher sample rates, that's sort of the definition, more samples per second. And what does phase have to do with it? Multiple sounds can't be phase aligned. The same sound from multiple places could in theory be phase aligned but not different sounds.

 

[DavidNJ]

As immutable once they are captures as aliasing they are there.

As to phase aligned imagine the signal 100100100100 and 010010010010 sampled at 3 times (3,3,2), 4 times(2,2,2,2), 6 times (211211), or 12 times (110110110110). Both have a frequency that individually are fully sampled at 4 pairs, but are not accurately represented. Equivalent to a loss of contrast, reduced MTF, in the image.

 

[Noiz2]

As immutable once they are captures as aliasing they are there.

As to phase aligned imagine the signal 100100100100 and 010010010010 sampled at 3 times (3,3,2), 4 times(2,2,2,2), 6 times (211211), or 12 times (110110110110). Both have a frequency that individually are fully sampled at 4 pairs, but are not accurately represented. Equivalent to a loss of contrast, reduced MTF, in the image.

???????

I think your peaking man. Stay away from ledges;~)

 

[DavidNJ]

No, just showing that oversampling, even if it down sampled for release,captures more relevant information in the frequencies (audio, spatial) being sampled.

 

[Noiz2]

No, just showing that oversampling, even if it down sampled for release,captures more relevant information in the frequencies (audio, spatial) being sampled.

This is getting way to narrow for a forum thread. But, there has never been a question about whether higher sample rates capture more information. The issue is the "relevant" portion. And that has a lot to do with work flow and end usage. Answering machines would capture "more information" if they recorded everything at 196k, but would it be relevant, I would argue no. Should you record the local death metal band at 196 and before the amps? Seems kind of pointless after they have gone to so much trouble to trash the sound. You would get a "better" recording though. If your recording acoustic music then I might slide toward the high end, but only if I have mics that can provide a decent signal up there. Recording with a dynamic mic is probably a wank because the physical mass of the voice coil is going to keep it from transducing anything useful in the high freq. range. You might get a touch more smoothness in the mid? maybe. But in the end it comes down to what people can hear. If your an academic then go for it but if your working in the real world you need to do a cost benefit analysis. High SR mostly loose in that case. If your recording for film and your not recording something that is going to be heavily manipulated in post then NOBODY is going to hear any benefits you might have gotten from recording at a high SR. And all of those recordings are going to get down sampled to 48 in editorial. Whether you get any down sample benefit depends a LOT on how it's done. Some conversions are brutal on the audio (QuickTime being one) and will come out worse than if you had recorded at a working rate in the first place. In the end those recordings are going to be mixed with Production audio at 48, CD sound libraries (originating at 44.1/16) and libraries of FX that have been recorded on everything from 1/4" tape to cass. to Mini Disk to DAT. Sample rates (original) will be all over the map, some sounds will have been processed to death (maybe). The point is it's a stew and it's a stew often made on a short time line under a lot of pressure and these days often by an understaffed crew. Other than taking the time to curse out and make fun of the "idiot" that recorded car bys at 96 that "we" now have to take the time to convert, they are not going to take a lot of time with your recordings. So if they have a decent batch converter maybe you get lucky, if not then some assistant or intern is going to churn them all through Audio Hub or Switch and nobody is even going to listen to them till they are at 48.

So that is sort of a long way to say that in post 99% of the time the higher SR has no "relevant" benefit.