Juno-G Sample/Output Filtering Comparison

Forum for JUNO-G
User avatar
dbijoux
Posts: 1700
Joined: 22:49, 26 April 2009

Juno-G Sample/Output Filtering Comparison

Post by dbijoux »

I was doing some comparison testing with my cables. After getting positive results, I decided to take it a bit further and make something a bit more useful to share.

What you'll see in the graphic below is that the Juno-G(and all Fantom Engines for that matter) actually filter the output of samples. It's came up in quite a few threads that Roland adds preemphasis to their internal samples. If you are using external samples, or resampling, preemphasis is required if you want to hear the same sound using the audio outs.

Image

The conclusion I've come to is that the output filter is very close to a 6db lowpass @ 12khz. It's slightly less nearing 20khz and drops off progressively more above that, but I don't have a filter less than 6db to test.

The interesting, and slightly troubling, thing I discovered is that adding preemphasis, while virtually restoring the frequencies perfectly(7-10k), does inadvertently create a high shelf almost 5db higher @ 20khz when compared to the original. It most likely would be EQ'd out in production, but without that and multiple samples could add up to quite a sheen in those extreme frequencies.

The filtering below 30hz can also be noticed, but is mostly beneficial.

(Note: These tests are certainly not the most scientific. I did not go to the extent of precisely aligning levels and the graphic is a composite of many takes, but the general idea should be clear enough.)
User avatar
dbijoux
Posts: 1700
Joined: 22:49, 26 April 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by dbijoux »

Upon further investigation, and bringing up a bigger concern with using the preemphasis method on externally recorded samples, I just noticed a noise floor increase of nearly 30-40db when the emphasis 'kicked into' the higher frequencies. I am not sure why the noise floor was raised so improportionally to the signal, but it certainly is high enough to be in the audible range (-60db) and could be problematic.

Hopefully I haven't rambled on too long now and this is still useful.
User avatar
Artemiy
Site Admin
Posts: 19754
Joined: 13:00, 17 April 2003
Location: Ukraine
Contact:

Re: Juno-G Sample/Output Filtering Comparison

Post by Artemiy »

My ears never told me there is a problem with this scheme, and all works fine for me with pre-emphasis. I'd use a Roland instrument for playing samples because it adds some character to them, and most software samplers play it exactly the same with no difference.
User avatar
dbijoux
Posts: 1700
Joined: 22:49, 26 April 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by dbijoux »

(UPDATED)
User avatar
dbijoux
Posts: 1700
Joined: 22:49, 26 April 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by dbijoux »

Ok, more testing and I see the source of my confusion.

What had initially made less sense to me was that the noise floor bounced not in proportion to the emphasis gain. After a closer look I see that it's more likely some DSP artifact from the PreEmphasis when the frequencies rise above 20khz!

Obviously, this is just a bit out of spec and quickly fading as any concern. While I thought it was a good idea to use the 30khz sweep to see what else might be going on, it ended up with me jumping to a conclusion.

Now back to your regular broadcasting. Thanks.
User avatar
Andy Keys
Posts: 1843
Joined: 13:44, 7 August 2007
Location: UK

Re: Juno-G Sample/Output Filtering Comparison

Post by Andy Keys »

Geek! :-)
User avatar
dbijoux
Posts: 1700
Joined: 22:49, 26 April 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by dbijoux »

Posiblemente!

It was worth a little effort tho. The real project lowered a 60hz bump in the floor over 5db + a few extra side efffects.
TheMvuser
Posts: 16
Joined: 18:08, 24 January 2010

Re: Juno-G Sample/Output Filtering Comparison

Post by TheMvuser »

Hi dbijoux. I just found this.. nice work.

most people think that those hi frequencies don't matter because they don't hear them.
well, in many cases they aren't listening closely enough.. those hi frequencies are what your drum transients live on. a transient is fast and fast = hi frequencies.

proper handling of this emphasis is why one persons track work will sound better and have more impact than another.

you can always de-emphasize to roll out those transients but unless you pre-emphasize you won't even have them.

it is better to keep around at least -13db in sample peaks for this reason. if you also look at what the industry pros are currently talking about, they are also mentioning that -20db in digital is what is really equivalent to real analog signal levels.

feed a digital compressor at -20 in digital and it sounds like it is supposed to.
because the compressor is modeled on analogue gain stage structure.

another good reason to hang at -13db or below.
lawndart
Posts: 35
Joined: 05:59, 14 August 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by lawndart »

Actually, it's probably not the output of the samples that is filtered - it's the input. Sample input has to be filtered before the input is fed to the analog-to-digital converter. If there was no filter before the ADC, any high frequency components above the Nyquist frequency (sampling rate/2) would alias, manifesting as noise spread across the entire audible frequency range in the sample.
User avatar
dbijoux
Posts: 1700
Joined: 22:49, 26 April 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by dbijoux »

Thanks MV, glad you approve, since you are the only other person I've heard mention a frequency sweep test. It's really not such a geeky thing, I beleive it's rather important to the sound and overall quality of your recordings. At the very least, it's better to know a little about what's actually going on.

You're correct on the -20-13db levels. This is particularly related to recording in 24-bit where your levels don't need to be nearly as hot to maintain a dynamic range comparable to 16-bit. The extra headroom doesn't hurt for effects or post-processing either and since noise is far less a concern in the digital realm, there's really no need to track much hotter than even -20db.

Lawndart - Yes, it's certainly not a filter post ADC. It's my assumption there is an additional processing stage somewhere before the converters in the output circuit. Here's an old drawing I came up with(although that was to highlight the filter comes before resampling).

http://forums.rolandclan.com/?action=show_thread&thread...

The filters in question are specifically related to a Roland engineering decision to preemphasize internal samples. A tactic commonly used in DSP to minimize noise throughout the signal chain.
User avatar
dbijoux
Posts: 1700
Joined: 22:49, 26 April 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by dbijoux »

I just read that again, I actually meant DAC. :)

This really has nothing to do with the Analog to Digital converters as we are talking about samples already in the digital realm(whether they were recorded on board or transferred from the computer). Of course, you were right about filters on the ADC too, but there's more to this story.
TheMvuser
Posts: 16
Joined: 18:08, 24 January 2010

Re: Juno-G Sample/Output Filtering Comparison

Post by TheMvuser »

dbijoux .


Roland often use the AK4528 Chip. I have the spec sheets for this and it looks like the pre-emphasis pins are pins 20 & 21. This is what they use for the MV8000 and MMP2 for example.
It's easier to find the spec sheets and see the filter characteristics in the PDF than write them here.

I am not sure if Roland nowadays always use the Chips own architecture to always apply or remove emphasis.

The Filters ADC Lo Pass is basically -6db @ 22.5khz.
The Filters ADC hi Pass is basically -3db @ 0.9 Hz and 0.1db @ 6Hz.

The Filters DAC has two pass bands.
Lo Pass is band one -0.01db @ 20khz.
Lo Pass is band two -6db @ 22.5khz.

The ADC LPF also has 4 pass bands but I couldn't be bothered to type it all :D

most people would think that those 0.01db values are totally insignificant but that is because they don't usually understand how filters operate. why should they ?
The 0.01db acts to shape the bend of the other -6db Filter slope.

not sure but, I think this is the frequency response of the ADAC in general and not the actual pre-emphasis curve. I think that might be separate.

The spec sheet says that the emphasis systems are switched in by control register (but) always OFF at double rate... e.g. OFF when running at 96kHz.. which the MV8000 doesn't run at. so it is dependent on the MV OS switching the flag to be on or off.

overall it doesn't look like they say what the pre-emphasis filter curve actually is but I think it is in addition to the normal filter characteristics of the ADAC itself.. so that could explain why my trace showed a 12db increase in amplitude of frequencies at the 20kHz point.

btw, it isn't easy to get exactly what the frequency trace of this curve is by using simple EQ boosts or cuts.
it's more complicated a curve than that. the easiest way is to have a good match EQ track what the differences are in a -12db white noise signal and the same signal with the Roland pre-emphasis applied.
Logics match EQ can do this spot on for example.
When you do a null test with phase inversion it cancels out to -90db etc.

overall, the main problem is that Rolands do not have gain change algorithms.
so you have a problem if you are at normalized levels.

hope this is of some value.

I urge people to ask Roland to add Gain change algorithms.
TheMvuser
Posts: 16
Joined: 18:08, 24 January 2010

Re: Juno-G Sample/Output Filtering Comparison

Post by TheMvuser »

lawndart Hi.

Quote:
Actually, it's probably not the output of the samples that is filtered - it's the input. Sample input has to be filtered before the input is fed to the analog-to-digital converter. If there was no filter before the ADC, any high frequency components above the Nyquist frequency (sampling rate/2) would alias, manifesting as noise spread across the entire audible frequency range in the sample.

I think I may have addressed this in my last post. the Filter responses for Nyquist are handled by the ADC chip in it's own architecture. pre-emphasis and how Roland use it is totally different. It is more like the RIAA lathe cutting curve idea. if you plug your record deck into normal line inputs and not the ones for the deck, you will then hear what is (really) recorded on the Disk. the Vinyl cutting lathe takes out the lows by a filter and when you put that into a phono preamp the inverse curve is applied.

This is why some people pay a lot for some phono preamps. it makes the record sound better because the quality of the reconstruction curve is better.

Rolands emphasis system is similar to this. It helps reduce the noise floor in Digital because the noise is usually heard at hi frequencies. so boosting them by 6 or 12 gives you a better SNR.

The upshot of this though is that the (output) signal you are listening to on Rolands is often, subduing the hi frequencies and therefore transients are less intelligible unless you apply the inverse boost of that Cut to the actual samples.

so to sum up, the outputs DO in fact cut highs all the time on a number of Rolands.
The MV8000 and MV8800 for certain. I am not sure about the V-Synth I haven't tested that yet. but I would like to know.

hope this clears a few Qs up.
lawndart
Posts: 35
Joined: 05:59, 14 August 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by lawndart »

MV, we're in agreement. What I originally thought dbijoux was comparing was:

signal direct from the function generator to the spectrum analyzer
vs.
signal sampled via the Juno-G's inputs, then played back through Juno-G outputs to the spectrum analyzer

What he actually did was import the digital signal as a .wav file downloaded from the computer. Therefore, what I wrote doesn't apply because the analog inputs were not used.
User avatar
dbijoux
Posts: 1700
Joined: 22:49, 26 April 2009

Re: Juno-G Sample/Output Filtering Comparison

Post by dbijoux »

Not much to add to this. It wasn't the most scientific experiment, but the graph highlights the basic assumption. MV's explanation also adds quite a bit to the discussion.

Pardon any misunderstanding on the setup. I may revisit sometime to look further into resampling and the deemphasis/preemphasis curves. Lawndart brings up an idea - to work out some specs with the inputs as well.
Post Reply