Our new mixer prototype described at https://ez.analog.com/message/165262#165262 suffers from noticeable output noise with its inputs turned down. To trace its source I modified the SigmaStudio project to switch a full-scale 1 KHz tone directly into the relevant output DAC. By measuring the level at the mixer's balanced output with this tone switched ON and OFF, we originally found only a 65 dB S/N. Then we discovered that our millivolt meter's 300 KHz bandwidth was allowing out-of-band noise into the measurement. Attempts to limit this bandwidth by RC filter at the meter input proved unsuccessful, so my friend hit on the idea of using a handy ADAU1452MINIZ board loaded with a modified version of the project at https://ez.analog.com/message/147007#147007 -- we have to be creative where I work, as there won't be a AP on the bench anytime soon. Reading the -1452's on-screen level meter yielded a more respectable S/N of about 80 dB, yet this remains below our expectation.
Running the same test on a ADAU1701MINIZ board shows a much better S/N of 96 dB at the board's headphone jack. Thus the question, what is making this mixer prototype so much noisier? Perhaps those with experience with noise could help us find out why. Here's what I know so far:
Audio stages run on 24 V with a 12 V pseudo-ground provided by a "Rail Splitter" IC. 3.3 V supply for the SigmaDSPs is a buck converter from the 24 V. The board is four-layer with generous ground planes and bypass caps. A passive output reconstruction filter is shown above (we're switching to active in the next proto). I don't remember the IC numbers although the op-amp is a low-noise type we've been happy with in other products, and the balanced driver is a ADI part.
Any suggestions on how to quiet this thing down, or how to track the cause of the excess noise, would be appreciated.
All I do are split-supplies, usually fed by a switcher running off of a battery. As for the cap value, it's not just a matter of stability, it's the amount you need to get the supply quiet. If…
My #1 suspicion would be the split-rail supply. You don't say which IC you are using, but TLE2426 for example has both standard and noise-reduction versions (the standard spec is 120uVRMS). Whenever…
I have dealt with this sort of problem a lot in recent years with trying to design single-ended circuits to work using a split rail power source. I concluded that it is difficult to do! Kind…
My #1 suspicion would be the split-rail supply. You don't say which IC you are using, but TLE2426 for example has both standard and noise-reduction versions (the standard spec is 120uVRMS). Whenever you are running a switcher on a PCB, it's tough to keep that noise completely out of the power supply. So either rail-splitter version will require ample bypassing. The opamp and line driver will have good PSRR, but any noise that is present on the virtual ground will directly couple into audio.
There is a spec in the SSM2142 datasheet that bothers me (I have no experience with that part), which is:
SNR ref 0V -93.4dBu
Headroom clip level 10.4VRMS 93.4dBu
Clip level is thus +22dBu, so I am uncertain if its noise floor is -93dBu or -71dBu. Those aren't very clearly written specs. If the latter, that could be your issue.
R235/R234 are higher than I like to use, but given the noise you are measuring (and the DAC output performance) I don't think those are the problem.
I would remove the ADAUs from the equation entirely, terminate VOICE MIX to ground, and do my noise analysis from there. You need to get happy with the analog performance of the circuit, first with the switcher off, then maybe with a dummy load on the 3.3V to simulate the ADAU, then switch on the oscillator, etc. Then finally reconnect to the ADAU.
For cheap audio frequency analysis, I'd look at picking up an inexpensive USB interface such as a Focusrite 2i2, and use any of the cheap or free software analysis tools. Nothing against -1452, but it won't give you the FFT resolution and analysis tools you can get on any PC.
I am heartfelt impressed and encouraged by your prompt, detailed, and helpful answer to my request for help.
Yes, the SSM2142 Data sheet's noise spec is quite confusing -- according to Wikipedia, dBu means "dBm unloaded", i.e. 1mW @ 600 ohms = .775V. Headroom below clipping should be, as you noted, a much higher number -- and not in dBu since it's a straight V/V ratio. AD has a reputation for exquisite data sheets, I hope they can maintain it.
I too am worried about the rail splitter in this mixer, although we have plenty of cap at its bypass input, and enough output cap to stay well over the hump region where it may go unstable. Looking at the gain-of-six stage between the DAC and output driver, the op amp's NI input refers to the 12 V pseudo gnd -- if I understand this correctly, any noise on this "gnd" actually gets multiplied by six (yikes!). Our previous all-analog models had true split supplies.
I just tried your excellent idea of removing the -1701 from the measurement to trace through the problem. Lifting the DAC output and grounding the amp input there produced 85 dB S/N. Interestingly, shorting R135 did even better at 95 dB.
Doing low-budget audio analysis with a USB front end is another gem. My friend has a Edirol U25 that we'll press into this service as soon as we can get through a backlog of unrelated work.
Thanks so much,
All I do are split-supplies, usually fed by a switcher running off of a battery. As for the cap value, it's not just a matter of stability, it's the amount you need to get the supply quiet. If R135/234 & SSM2142 is the only use of +12V, you can add some series resistance, maybe as much as 47R, before its bypass cap to improve its filtering. I have settled on 10R/470uF for my basic circuit but if I had space I would use a larger cap. For the series resistor, SSM has specified input impedance of 10K, but since it doesn't have much gain you might be able to tolerate some DC offset. And the LF gain error at U18A as a result shouldn't be of too much worry.
I am not sure why shorting R135 helped, because its thermal noise should be shunted by grounding C240 (or connecting it to the DAC).
One thing I should have mentioned before was the grounding scheme, these are often trouble with switching supplies. What I have found best after repeated attempts to be fancy is a single continuous ground plane with the switcher as distant as possible from the audio circuit, with good local bypassing of audio circuits and all audio grounds referenced together as closely as possible. I found that works better than creating split planes for the PSU, but not everybody agrees with that and I have done "clean" and "dirty" grounds for customers that wanted that approach.
I have dealt with this sort of problem a lot in recent years with trying to design single-ended circuits to work using a split rail power source. I concluded that it is difficult to do! Kind of like mixing water and oil. If you look at the DAC outputs and think of the return current path you will see that it includes the pseudo ground generator. So that will add its noise to the signal. So looking at it the other way, the op-amp is comparing the signal to the ground + the pseudo ground generator so it will take the difference of these two signals.
My solution in the past was to treat the signal as a differential signal and setup the opamp in that manner.
I will show what I am thinking about.
This first circuit is a standard diff input with bi-polar supplies and referenced to a center ground. You can always drive this with a single ended signal by just tying the In - input to ground.
Now this circuit is with the op-amp running off of a single ended 24V power rail and a 12 V pseudo ground. Then tie your inputs to the DAC output and its reference which happens to be ground. Try this and see if it helps. I would also run this trace from the cap (C4) all the way close to the DAC and tie it to the ground plane there. This way you will be sensing ground as close to the same place as the DAC is.
I have used inamps in a similar manner to fix that problem, that also allows noise cancellation of the virtual ground since both inputs get referenced.