XS902 De-esser - First Prototype!

Hi all,

Following on from this thread, I present the first prototype of the XS902 De-esser:





..."XS902"..."X...S"..."Ex-Ess"...geddit? hyuck! :green: *sigh*...

The meter board just plugs into the main board via a 4-pin header, and sits on top of a pair of 25mm spacers.

Schem I'm working from is here:

http://i153.photobucket.com/albums/s222/ac1176/XS902_V0_schem_hires.gif

I have yet to draw the diagram for the meter board.

At the moment I'm having some trouble getting the unit to properly detect sibilance, but I think the problem is centred around the comparator/JFET section (IC6d, Q1 in the schematic) - it seems to be always "on" allowing the sidechain signal to pass through to the VCA at all times. For example, adjusting the "range" control with no signal input results in the VCA gain being affected, which definitely isn't right.

Any thoughts?

Edit: link to higher resolution schematic modified.

That's great progress, considering we had the DBX902 dissection discussion only a few weeks ago.

Have you tried the '902 calibration procedure, as listed in the 902 Service Info zip file on the dbx ftp site ?

Good luck,

JDB.

I'd be interested in seeing a legible version of the schematic... I've discussed making a 'clone' of the 902 with a couple of lab members.

The detection is indeed critical, are you getting a voltage swing at the "op-amp side" of the diode if you 'burst' the input with HF?

Keith

Interesting, I can't read the small print on the schematic either but it looks like a wide band VCA mode and a HPF only VCA mode. The side chain has both full range and HPF only results so it can respond to relative HF content or absolute. Can't tell any more than that with these old eyes. Looks like it could be pretty flexible.

It should not he much effort to add compression, limiting and other normal dynamics processing to the wide band path, since the money has already been spent on VCA and detection, but I guess this design may be panel space limited for controls.

JR

Correct on the analysis John, as is the 902.

However, the 202C VCA is no longer available, nor is the original RMS-DC converter.

The 902 is the most flexible VOCAL de-esser in existence, because it 'adapts' its threshold depending on overall vocal effort. If the vocalist 'whispers', then it lowers the de-essing threshold sensibly, and conveniently... it just works 'Right'.

Full-range music is a different matter, and the 902 isn't much used in these applications (even though it offers the 'HF-ONLY' GR mode) but for vocal overdubbing, or hi-hat-into-snare-mic bleed reduction, the 902 is a belter!

If I could read the values I might be able to help somewhat.

Cheers,

Keith

[quote author="SSLtech"]If I could read the values I might be able to help somewhat.[/quote]
Curtis' schematics (and part values) are pretty much a direct copy of the 902 schematics which can be downloaded from http://www.dbxpro.com/Download/discontinued.htm (direct links: schematics and service info).

The only changes that were implemented (per the previous discussion) are:
- op-amp output stage instead of the original discrete one
- zero-CV bypass instead of a direct bypass switch
- replacement of RMS/VGA hybrids by modern THAT parts
- adjustment of VGA control voltage from 50mV/dB (for the older VGA) to 6mV/dB

(As for that last point: Curtis, I see you've handled that by reducing R37. Considering noise and opamp offset voltages (and old-circuit compatibility) it may be better to stay 50mV/dB throughout the XS902 and only drop down to 6mV/dB right before the THAT218x. Oh, and are you sure that TL074 is a straight replacement for the original threshold comparator?)

JDB.

I have never used DBX rms chips myself, the one comp I'm aware of at my old day job that used them was done by a junior engineer who was able to get it done from the app notes without my participation.

When performing log domain math with opamps it's always good practice to keep voltages scaled up high enough to minimize noise, and opamp DC offset errors. Further using a passive resistive pad right at the VCA pins to scale down control voltage, will mitigate against layout issues too. Hum or low level noise on a VCA control line will modulate the signal with that noise.

Looking at that schematic there looks like a unity gain opamp buffer (with some spare parts) in series with control voltage line. This may be in error but as drawn, it is introducing at least the DC offset error of that opamp into the VCA control path which could be multiple dB for some non precision opamps. I routinely used TL074 for such log math but with the control voltages scaled up where small DC errors were inconsequential when later padded down.

JR

[quote author="JohnRoberts"]When performing log domain math with opamps it's always good practice to keep voltages scaled up high enough to minimize noise, and opamp DC offset errors. Further using a passive resistive pad right at the VCA pins to scale down control voltage, will mitigate against layout issues too.[/quote]
Agreed, however the THAT218x VCAs don't have an internal control voltage buffer and require a very low impedance drive ('zero source impedance', says the THAT datasheet). A plain passive pad would increase distortion; THAT recommend a circuit like the GSSL uses (low noise op-amp buffer with a resistive pad just before it).

[quote author="JohnRoberts"]Looking at that schematic there looks like a unity gain opamp buffer (with some spare parts) in series with control voltage line. This may be in error but as drawn, it is introducing at least the DC offset error of that opamp into the VCA control path which could be multiple dB for some non precision opamps.[/quote]
Yup. I assume you're talking about the op-amp just after the upper RMS converter, it's rigged as a -1 amplifier to subtract the log RMS of the high pass filtered portion of the signal from the log RMS of the full signal. There is an offset trimmer on the subsequent stage, but that op-amp had better not drift too much.

What the '902 does is conceptually very simple:

VCA CV = logRMS(full BW) - logRMS(HPF output) + threshold.

The de-esser starts reducing gain when the energy in the HF bands of the signal is {threshold} below the total signal energy; that way the de-essing does not depend on the absolute level of the signal. Maximum GR is equal to {threshold}; this happens when the signal has only HF content.

JDB.

Have you tried the '902 calibration procedure, as listed in the 902 Service Info zip file on the dbx ftp site ?

Click to expand...

Yep, but until I get this CV-bleedthrough problem licked I can't get very far into it.

I can't read the small print on the schematic either but it looks like a wide band VCA mode and a HPF only VCA mode.

Click to expand...

D'oh! I hadn't realised the diagram had shrunk that much. Sorry 'bout that, I'll do another export at a more decent resolution later on today.

Oh, and are you sure that TL074 is a straight replacement for the original threshold comparator?)

Click to expand...

Pretty sure. The original uses an LF353 which should be close enough for a TL07x to sub with. I suspect that the real problem is that the original RMS module has a different (larger?) output voltage swing than the 2252, and so the full BW-RMS'ed output runs at the wrong voltage levels to trip the comparator properly.

When performing log domain math with opamps it's always good practice to keep voltages scaled up high enough to minimize noise, and opamp DC offset errors. Further using a passive resistive pad right at the VCA pins to scale down control voltage, will mitigate against layout issues too. Hum or low level noise on a VCA control line will modulate the signal with that noise.

Click to expand...

Would I perhaps be better off scaling R37 back up to its original 247K (the feedback R on the precision rectifier section, currently 30K) and then scaling the CV back down again at IC6a just before the CV input to the VCA?

[quote author="jdbakker"][quote author="JohnRoberts"]When performing log domain math with opamps it's always good practice to keep voltages scaled up high enough to minimize noise, and opamp DC offset errors. Further using a passive resistive pad right at the VCA pins to scale down control voltage, will mitigate against layout issues too.[/quote]
Agreed, however the THAT218x VCAs don't have an internal control voltage buffer and require a very low impedance drive ('zero source impedance', says the THAT datasheet). A plain passive pad would increase distortion; THAT recommend a circuit like the GSSL uses (low noise op-amp buffer with a resistive pad just before it).

[quote author="JohnRoberts"]Looking at that schematic there looks like a unity gain opamp buffer (with some spare parts) in series with control voltage line. This may be in error but as drawn, it is introducing at least the DC offset error of that opamp into the VCA control path which could be multiple dB for some non precision opamps.[/quote]
Yup. I assume you're talking about the op-amp just after the upper RMS converter, it's rigged as a -1 amplifier to subtract the log RMS of the high pass filtered portion of the signal from the log RMS of the full signal. There is an offset trimmer on the subsequent stage, but that op-amp had better not drift too much.

What the '902 does is conceptually very simple:

VCA CV = logRMS(full BW) - logRMS(HPF output) + threshold.

The de-esser starts reducing gain when the energy in the HF bands of the signal is {threshold} below the total signal energy; that way the de-essing does not depend on the absolute level of the signal. Maximum GR is equal to {threshold}; this happens when the signal has only HF content.

JDB.[/quote]

No I was talking about the opamp right at the VCA. While I may be reading it wrong, it looks like the - input is actually connected to the VCA control port and the output not connected to anything but the - input trough a parallel R & C? As drawn a unity gain buffer with a strange source impedance characteristic, but I probably need to shut up until I can read a schematic I can see.

An opamp buffer is fine if opamp is low noise and has good DC characteristics. I vaguely recal some pretty cheap general purpose opamps used in those sockets on some old designs.

My recollection was control ports should be an low and the same termination impedance. If the other port is grounded then good opamp can get you close. I recall passive pads with similar (low impedance) resistive terminations at the other control port delivering acceptable THD.

Scaling values up is pretty straightforward. The rectifier can be scaled so output is large but well within PS limited voltage range for anticipated input peak level. The more you can divide it down later at the VCA the less layout, noise and DC errors will get into the VCA.

The log math is pretty straight forward but again I need to wait until I can see schematic better. I think I see some opamps with resistors in the + inputs to cancel input bias current errors that aren't needed with bifet opamps. No big deal mostly spare parts doing nothing. Original designs may have had 741 equivalents in those locations.

JR

No I was talking about the opamp right at the VCA. While I may be reading it wrong, it looks like the - input is actually connected to the VCA control port and the output not connected to anything but the - input trough a parallel R & C? As drawn a unity gain buffer with a strange source impedance characteristic, but I probably need to shut up until I can read a schematic I can see.

Click to expand...

No, you're reading that right, and that's how it is in the original too. It gives a handy point to break into the CV for things like stereo de-essing and other special sidechain functions (or at least that's how the manual describes it). I left it as-is as I intended it as an option for combining two units as a stereo pair.

New high-res schematic located here.

[quote author="Curtis"]

No I was talking about the opamp right at the VCA. While I may be reading it wrong, it looks like the - input is actually connected to the VCA control port and the output not connected to anything but the - input trough a parallel R & C? As drawn a unity gain buffer with a strange source impedance characteristic, but I probably need to shut up until I can read a schematic I can see.

Click to expand...

No, you're reading that right, and that's how it is in the original too. It gives a handy point to break into the CV for things like stereo de-essing and other special sidechain functions (or at least that's how the manual describes it). I left it as-is as I intended it as an option for combining two units as a stereo pair.

New high-res schematic located here.[/quote]

Thanks, I can now read the schematic better, and that opamp still looks unconventional to me, to put it kindly.

From the 2181 data sheet they advise keeping the drive impedance at control ports to 50 ohms or lower. Further they caution if using an opamp to buffer that point the source impedance of an opamp which rises at HF due to falling loop gain margin may cause instability. The recommend adding a R (100 ohm) in series with a C (1.5nF) from the control port to ground for stability.

Driving the control port (as drawn) from the minus input of the (tl074) opamp will be moderately low impedance as long as the opamp can keep up with current drawn at the node though the feedback resistor. That point will be higher source impedance than a typical opamp "output" like THAT warned against, and who knows what else is hanging off that - input to bother the opamp stability. The better node to use IMO for stereo de-essing or sidechain tricks is the + input, before it is buffered.

Another tidbit from the 2181 data sheet is that the noise of a unity gain connected 5532 will increase the the noise modulation floor from -94 dB to -92dB. OK only 2 dB but the 5532 is far quieter than a TL074 assuming it was connected right, not to mention TL074s unless selected or graded parts can have DC offsets of several mV perhaps significant at the control port of a VCA.

While I would have used a resistor divider with the final termination resistor < 50 ohms, I appreciate the opamp mentality and modest voltage available from the RMS chips. This is one application that screams for modern low noise, low DC offset super opamps but it seems a waste of $ to throw them in a control side chain. The old general purpose bipolar opamps were a somewhat better match for the task than older bifet technology.
------

Looking at the rms/log/control voltage circuits, It looks like the 2252 is designed to interface with the VCA 1:1 so is only hundreds of mV full scale.

That said, all the opamps in that area would benefit from low noise and low DC offsets, while there is a dedicated trimpot just for DC offset this circuit wasn't designed for bifet sized offsets. The 4.7k resistors to ground from the + inputs on several of the opamp sections are a holdover from an earlier design version that obviously used old school bipolar opamps, and the Rs were there to correct for input bias currents.

Since all of the control voltage crunching is done at 1:1 or close to it the layout is important. The ground end of R42 (12k) should be close to the VCA ground.

Without scaling up the control voltage I suspect the TL074 is not a great opamp for this application, something with better DC performance, lower noise, etc might be better. The opamp that really makes me uncomfortable is the one driving the VCA control voltage. If the VCA is unstable because of impedance there consider adding the RC they mention in the spec sheet.

Considering that PCB is already finished perhaps a better high performance quad opamp might drop in.. while I'd take my exacto to that one opamp, if it works and you're happy enjoy..

JR

Thanks for the input John. The design at this stage is still just a prototype, so revisions to the PCB layout isn't necessarily a huge issue. I've already taken to the meter board with an exacto and wire bridges, so a rebuild of this PCB is on the cards. I wanted to adjust some of the component spacings on the main board anyway, so I'll be doing a tweak of the layout regardless.

I've got the comparator section working properly now - increasing R32 from 100K to 180K lowers the trip point of the comparator back down to -120mV, so with no/low signal input the comparator output forces Q1 on, shutting off the CV as expected. The unit de-esses now, but I've yet to examine it closer with a scope/sig gen/meter to find out if it's working properly.

I agree, the opamp arrangement at the VCA CV input is quite odd, but it does seem to work. I am however open to alternatives, so I might explore different arrangements with NE55xx's or similar

The way I see it, the op-amp is always going to 'fight' to equalise the two input voltages, and so the R/C combination is going to cause the op-amps output to swing about pretty entertainingly while forcing the inverting input node to match the noninverting input.

Simply connecting the output through NO resistor back to the input would also give you the same result at the input node... UNTIL you connected it to a second input node ("stereo link").

Viewed as-drawn, it looks needless, but you can't connect two op-amp ouputs together through zero ohms without someone having to 'hold their coats', so I'm guessing that this is why there's some resistance... the cap is still a mystery to me, but more may suggest itself to me as I ponder it a little longer.

"...This is a three-pipe problem, Watson!"

Keith

As the data sheet suggests the rising output impedance of a normal opamp is undesirable, putting a resistor in series with the TL074 output which isn't known for it's output impedance before applying the magic of NF makes the impedance there even worse.

Yes that configuration allows the two points to be shorted together, but then two opamps will both be fighting each other trying to regulate that same point so summation will be nonlinear and one or both opamps will probably saturate (clip). When both clip the impedance at that point becomes hundreds of ohms.

I suspect any artifacts caused by that combining may be difficult to hear with complex music, but it is not good design IMO. Combining at the + input does not suffer those problems while even that opamp connected properly doesn't completely get the job done, adding the recommended RC will get the termination low at HF if sticking with an opamp buffer.

There are multiple ways to clean this up.. easiest might be to drop in a precision quad, a little harder is to scale up the voltage from RMS chips so opamps are less critical and divide down result at VCA. FWIW that circuit is already boosting and dividing about 3:1 at intermediate stage but I'd be inclined to go at least 10x or more to use Bifets (which I did routinely in similar circuits). Why limit the opamps to a few hundred mV?

JR

Yes I'd thought that -while the two inverting poles will become one- the two noninverting inputs will still be living seperate lives... but it reduces the fisticuffs between the two op-amps to a mere shouting match...

It still is -as you rightly observe- a VERY odd arrangement, and prone to error rather easily.

-And yet it's so unusual that it almost MUST be intentional...

I have a couple of 902s here; does anyone want me to take any specific measurements?

Keith

It strikes me as a mistake or two, that kind of worked so it wasn't fixed.

I have in fact used a somewhat similar opamp topology in a tape NR where I intentionally counted on an opamp clipping to change the impedance at it's - input and provide a threshold related time constant shift. Very fast attack for large transients, returning to slow and smooth filtering after transient decayed.

I don't see any incidental benefit from this topology in the de-esser, only downside. Combining will be dominated by only one or the other in control. The one closer to 0V will have a slight advantage so when linked operation will follow the less sibilant channel, or no de-essing if both opamps clip when the control voltage will settle to roughly 0V or unity..

Intermittently driving the control port with several hundred ohms may cause audible artifacts that are hard to hear in the presence of sibilance.

JR

[quote author="JohnRoberts"]Intermittently driving the control port with several hundred ohms may cause audible artifacts that are hard to hear in the presence of sibilance.[/quote]
A fascinating observation.

-I do wonder if that may indeed be another reason why these work well on vocals (where 'unvoiced sibilants' as opposed to 'voiced sibilants' and unvoiced fricatives as opposed to voiced fricatives, are to be stomped-on in a full-band manner... 'Sss' as opposed to 'Zzz', 'Fff' as opposed to 'Vvv' and so on) but don't work anywhere near as well in HF-only mode on full-range program music...

I'm definitely MOST interested in where this is heading...

Keith

I doubt the intermittent THD would be very audible on top of sibbilants, but the data sheets suggest instability if that port isn't terminated < 50 ohms. So what is more likely to occur that "might" be audible could be a burst of oscillation in the VCA.

If this was occurring it should show up in testing with tone bursts, but only in stereo linked mode driven with different specific specta and level signals. I'm guessing this is stable for mono, despite the spec sheet warning as such would surely have shown up in general use.

My personal opinions may just be conservative engineering. The relative vs. absolute de-essing is interesting but as I recall the driving force behind developing de-essers in the first place was to deal with sibbilants, elevated by compression and close micing above their natural levels, overloading broadcast and recording HF pre-emphasis curves. At lower levels more sibbilant content is generally perceived as better due to good old Fletcher Munson losses.

I notice this de-esser stops working below a threshold which somewhat addresses my concern. The adjustable filter frequency makes this more than a simple de-esser and more of a spectral shaper. Probably useful in it's own way.

JR

Try this as an alternative (just a partial section of the diagram):

http://i153.photobucket.com/albums/s222/ac1176/XS902_partial_schem_rev1.gif

Precision rectifier feedback R increased to 100K giving 10x amplification to the RMS signals, voltage divider on output of rectifier (R41, R42) changed to give 1/10 division, CV voltage taken from output of the final buffer rather than the inverting terminal, series RC filter added to control port (Rxx, Cxx), sidechain opamps changed to NE5532's.

Physically the changes shouldn't be too hard to implement onto the existing layout.

Thoughts?