Low Noise high gain amp based on 18W kit: How and why

[nickb]

You are putting noise resistances, which are proportional to power, into an equation which needs voltages.

Yes, indeed I am - thx

I also fixed a typo above where I used "/" instead of "."

[SoulFetish]

I finally got around to reading through this thread. I checked quickly it out a few days ago and when I saw how the differential front end was set up, I wanted to wait until I could take some time to look over the rest of the circuit and think about it. It's been haunting me since I read things like "solving the instability of miller capacitance..."
I'm a sucker for creative ideas (both artistic and engineering), so I respect the project and think its a refreshing approach.
For now, I'll let the heavy weights fight over the which noise source dominates the audio spectrum in these operating conditions.
As for AES E-Library » Noise in Triodes with Particular Reference to Phono Preamplifiers... until there are changes in my own operating conditions, I'm limited to books and free technical papers.

[Mike Sulzer]

I have the measurements to show that flicker (1/f) noise is not the dominant noise source in my amplifier. (A phono preamp might be a different story.)

The technique is the following:

1. Connect a recording interface to the preamp out.
2. Turn gain 1 and gain 2 to 10.
3. Measure the spectrum (Electroacoustics toolbox) under two conditions:
____a. with the guitar input shorted. (The noise measurement)
____b. with a 10K metal film resistor, bypassed by a 47pf capacitor, across the input. (The noise plus cal measurement.)
4. Accumulate data in each of these modes for about 10 minutes. Spectra from dc to 24000Hz have 32769 points.
5. Process the data as described below.

Possible problems with the use of a resistor as a calibration source:
1. Excess noise from the resistor.
2. Noise from grid current.
3.Different frequency responses in the two measurements.

Making some measurements with a 4.4K resistor shows very nearly the same results and so these effects appear to be small.

Processing:
1. Read the spectra into Python using standard software for the Matlab format.
2. Convert from db back to linear power.
3. Filter the noise and npc spectra with a 51 wide median filter. (Removes power line and other spikes.)
4. Find the noise resistance as a function of frequency using this equation: Rn = Rc*noise/(npc - noise). Rc is the cal resistor, 10K.
5. Construct a model representing the sum of flicker and shot noise.
6. Use nonlinear least squares fitting to find the relevant parameters.

Nr measurements from 5 tubes (all Slovakian 12AX7) tried in the first stage are shown here:
The first three are the tubes that came with the kit. (blue, green, red) They are very consistent at the high frequencies and differ some in the low frequencies. The other two are tubes from another project. One of them is slightly worse at the high frequencies, but a bit better at the low frequencies. The last tube is a disaster and should not be used.

To get a somewhat more stable measurement that goes a bit lower in frequency, I averaged the noise and npc spectra from five of these measurements using the first sample tube. The result is shown here:
Also shown is the model using the results of the fitting. The fitted parameters are: shot noise level: 2982 ohms, b (flicker nose decay): .901, flicker power constant (K*i**2): 8.25e5. The first parameter is about what you would expect from a stage with the noise of two tubes.

Using the measurement for Rn, and a flat spectrum constructed from the shot noise level, the increase due to flicker is .58db. (Remember, the log frequency scale might appear to overemphasize the importance of the low frequencies.) This is a uniformly weighted result from 100 to 20,000 Hz. Since human hearing sensitivity is such a strong function of frequency, the true effect of noise needs to be a weighted measurement. This is next.

[Mike Sulzer]

When making engineering decisions which affect the noise performance in an amplifier, it is the effect of the noise on the ear-brain system that counts. Determining the effective noise is important.


The influence of noise in human hearing has been studied extensively. Relevant measurements have been made for over 80 years, and significant revisions have occurred as little as a decade ago. It has always been known that an effective number for noise power is obtained by weighting the noise in different frequency ranges, not by adding in an unweighted manner. However, the exact weights have changed over the decades. Weighting is required because of the huge variation in the sensitivity of human hearing as a function of frequency. In addition it is necessary to account for the particular characteristics of the equipment under design. For example, a phono preamp, which boosts bass with respect to 1KHz by 20 db, is very different from a guitar amp, which might cut bass.


There are various curves used in weighting noise. Three are found on this plot: , which can be found in various places, including here:
http://www.lindos.co.uk/cgi-bin/Flex...VIEW=full&id=2 .


The curve labeled A is based on the original Fletcher Munsom curves, which have been revised several times. The curve labeled 468 is for use with clicks and noise bursts. The one labeled (inverse) ISO 226 is the inverse on the 40 phon curve of the 2003 ISO 226 curves, the sucessor to Fletcher Munson. 40 phon is a quiet sound; noise weighting is based on this curve because noise should be quiet. Electric guitar is played so loud that the noise might exceed the 40 phon level by quite a bit, suggesting the use of the inverse 60 phon curve; the difference is primarily that the higher phon curves cut less bass. However, the amp and speaker cut quite a bit of bass, and so use the 40 phon curve and forget about the amp and speaker.


Although log-log plots show details over wide dynamic and frequency ranges, they are misleading. This plot uses linear-linear axes: . It is clear that the high power 1/f noise at the low frequencies is not a very large part of the total. The inverse ISO curve plotted in this way shows how little the high frequency noise contributes to what you hear even if you have good hearing. It is remarkable how much this high frequency cut is at about the same frequency as a guitar speaker. (Perhaps my parents and grand parents were on to something when they said that all they heard in rock and roll was noise.)


In the product, the bass still falls despite the 1/f effect. But the high frequency has been cut so much, that it might still count for more than in the uniform summation performed in my previous post. Yes, when the appropriate summations are performed, Rn is .82db higher than if the noise were flat. This is up a bit from .58db.

[nickb]

Nice analysis Mike- you've been busy.

I didn't quite follow what the equation for the fitted model was from your description - could you spell that out for me please?

0.82dB means the effective ENR is 3601 ohms over this bandwidth.

Merlin's paper experientially measures noise for various tubes. It is some comfort to see the data is very similar although the technique is a little different. Incidentally, the paper doesn't really have much to do with phono pre-amps. I suppose he had to come up with a catchy in-vogue title.

I suggest the log vs linear comparison is why two people looking a the same data reach two different opinions as to the contribution of flicker noise to the overall picture.

Since you've got yourself all set up with Python and the audio module, I think it would be really interesting to generate the audio for the shot only, shot + unweighted flicker +shot and finally weighted shot + flicker cases. That is something anyone could understand.

Thanks for the interesting & original work.

[Mike Sulzer]

Nice analysis Mike- you've been busy.

I didn't quite follow what the equation for the fitted model was from your description - could you spell that out for me please?

0.82dB means the effective ENR is 3601 ohms over this bandwidth.

Merlin's paper experientially measures noise for various tubes. It is some comfort to see the data is very similar although the technique is a little different. Incidentally, the paper doesn't really have much to do with phono pre-amps. I suppose he had to come up with a catchy in-vogue title.

I suggest the log vs linear comparison is why two people looking a the same data reach two different opinions as to the contribution of flicker noise to the overall picture.

Since you've got yourself all set up with Python and the audio module, I think it would be really interesting to generate the audio for the shot only, shot + unweighted flicker +shot and finally weighted shot + flicker cases. That is something anyone could understand.

Thanks for the interesting & original work.

Thanks, Nick. Here is the Python function that makes the fitted model.

# Function to make the spectrum of shot noise + flicker noise
# f: an array containing the frequencies
# The other three arguments are the parameters (single numbers)
# K: the scale constant for the flicker noise
# b: the exponent of flicker noise fall off (the color of the noise)
# sn: the level of the shot noise
def tn(f, K, b, sn):
return K/(f**b) + sn

The fitting routine then finds the values of the three parameters that result in a function that differs from the data the least, that is in a least squares sense. The result for shot noise includes that from gird current as well, but that should be small.

I will make some short aiff files containing white noise and the noise of a 12AX7. There are then several ways to go from there. I think one interesting thing to do would be to inverse weight both of them. That would show you how noise would sound if a healthy ear brain had uniform sensitivity across the band.

[Mike Sulzer]

Thanks, Nick. Here is the Python function that makes the fitted model.

# Function to make the spectrum of shot noise + flicker noise
# f: an array containing the frequencies
# The other three arguments are the parameters (single numbers)
# K: the scale constant for the flicker noise
# b: the exponent of flicker noise fall off (the color of the noise)
# sn: the level of the shot noise
def tn(f, K, b, sn):
return K/(f**b) + sn

The fitting routine then finds the values of the three parameters that result in a function that differs from the data the least, that is in a least squares sense. The result for shot noise includes that from gird current as well, but that should be small.

I will make some short aiff files containing white noise and the noise of a 12AX7. There are then several ways to go from there. I think one interesting thing to do would be to inverse weight both of them. That would show you how noise would sound if a healthy ear brain had uniform sensitivity across the band.

I have made three aiff files, each lasting about 5 seconds. They are the flicker noise alone, the shot noise alone, and the two combined, that is, the total noise. The three parameters derived earlier and the above function were used. First, the function was used to make the total model spectrum. Then the level sn was subtracted to give the spectrum of the flicker noise alone. The shot noise spectrum is just the level ns, independent of frequency, of course. Next, two random normal samples were generated using the Pyhon (numpy) routine numpy.random.normal. These were generated in the frequency domain. (That is, they correspond to the fft of the time domain random signal.) These are voltage, not power, and so they are scaled as a function of frequency by the square root of the spectral shape. The third random sample was generated from the sum of the two. It was then verified that its power spectrum matches the model spectrum generated from the function. This assures that the first two samples have the correct relative power levels, and that the overall spectral shape is correct.

The three random signal samples are then inverse ffted to get the time domain samples. They are scaled by a factor common to them all (to get them to fill properly into 2 byte integers with the correct relative powers), converted to 2 byte integers, and written out as .aiff files. These files are then zipped. They are best played through a guitar amp not too loud. I drag the files into MPLAY, making a primitive playlist and then loop them. I hear the shot noise as quite a bit louder than the flicker noise, and the total as just a bit louder than the shot noise and somewhat different in sound.
(The relative sizes of the zipped files indicate that the power levels are different.)

tbflicker.aiff.zip
tbshot.aiff.zip
tbtotal.aiff.zip

[nickb]

Interesting - thanks. I have my opinion. I wonder what others think?

Here's all three as one file. Shot then flicker then both...

[SoulFetish]

I've been wanting to circle back to this thread for a while. I have some thoughts on the input stage, but I kind of wanted to expand the conversation grounded plate/grounded grid stage designs in general. I have a thing for interesting and different approaches to stage design, so compound stages, grounded grid/plate, differential, and balanced stages hit my geek spot. (for instance, have no idea what a Mu follow is going to sound like in a guitar amp.... but I'm going to find out).
But, I do want to talk about some what I think are interesting points in this design and some question as well. So, should I branch off to a new thread? Or, keep this thread going?

[KevinOConnor]

Hi Guys

Overall I like this design. Well done Mike!

As you mentioned initially, the stage sequence is imperfect and to me there should be decoupling for each stage.

Split-rail preamps and amps using tubes are more commonly found in broadcast applications, such as microphone preamps and utility gain blocks of olde - I've built many. There are many advantages and many potential circuit options with the split rails, so for anyone wishing to build something aoing the lines Mike has shown, there are more things to try.

Each gain stage has essentially no gain at all until you add the cathode cap to ground. This cap can be sized to provide gain above whatever frequency you choose. In most of Mike's stages the gain is broadband like a Fender stage. To limit the gain, simply place a resistor or pot in series with the cap - with a pot you have a tone control that boosts. You can also simply place a pot in series with the large Ck shown and have variable gain for each stage in addition to the interstage level controls that allow volume to go to zero. If you had only the variable gain controls, gain to go to 0dB, which is unity or '1', but that would not set the volume to zero.

For really focused gain, an LRC series circuit allows the C to set the low-f rolloff, the L to set the high-f rolloff and the R to set the gain in between. There is a limit as to how many such elements can be tied to a given gain stage, but it can certainly provide a lot of flexibility.

It looks to me as if the gain is fairly flat frequency-wise as Mike has set it up. The voicing comes from using 10nF coupling caps and actually in the choice of EL-84s. I don't really care for EL-84s and were this my amp i would use octals, which allows mixing different types and/or using different types for many textures. I'm glad to see a correct screen-stop value used here!

Concertina splitters have always been known for poor supply noise rejection. The easy fix is simply to use a large filter on that node. In many applications, the concertina is early in the chain, followed by more gin before the power tubes, so it is a few supply nodes away from the output stage and thus well-filtered. In my amps, the concertina is placed right where Mike's is driving the output tubes, but my supplies are extremely well-filtered so no noise issue.

One option to fix the assignment sequence for the triodes would be to use the "weird" 12DW7 as V3. This would allow Xs for the first two, with the low-mu side of the DW used to drive the pedal and the high-mu side as the splitter. The 12DW7 has one section like a 12AX7 and one like a 12AU7. Yes, the plate drive impedance to the EQ would be increased this way, but my experience is that any plate-driven tone stack sounds better than a cathode-driven one.

Anyway, it looks like a great amp and has lots of potential for expansion.

Have fun

[Mike Sulzer]

Thanks; I will next make an octal amp using a similar design. The EL84 amp in a large chassis was an easy build for trying out ideas, etc., as well as a really great amp in its own right. (Using grid bias, IMO, makes the EL84 a better device, but that really depends on the sound you want.)

Also I will investigate the stability/noise trade off again. I think merlinb is correct that you do not need a grid stop resistor on a 12AX7 stage, at least considered alone, but instability in an amp can be feedback from many stages later, and a grid stop resistor can affect the stability in this case. So I will try the two sections of V1 in parallel for a very low noise design requiring very good lay out and shielding, but I will build it in such a way that it will be possible to change to the input stage of this amp also.

[KevinOConnor]

Hi Guys

There are a couple of ways to control the turn-on of the split-rail tubes so that Vhk is not exceeded: one uses simple diode clamps the other uses relay contacts.

A good example of the diode clamp is in Fender's Bassman Pro 300. The input stage of the PA is a split-rail-powered diff. Rk is tapped and a diode ties the tap to -16V. The diode only turns on during the start-up. Once the tubes conduct the tap point is pulled above the diode reference voltage and the diode turns off. This would work fine on your input stage but might interfere with signal handling on the subsequent stages unless a much lower negative reference is used for the clamp.

The relay clamp uses a relay contact that shorts the cathode to ground only during the tube warm-up period. This can be refined so the short is not quite a short but rather a kilohm or two. This method can be applied to all the stages including the split-rail concertina and works very well. We used both methods in broadcast microphone preamps.In those, there was already an aux supply for the relays and a turn-on control (TOC) circuit was already in place for output muting at turn-on/off. The TOC is just one BJT with another to handle the relay coils.

Some observers might not have noticed that the signal level in Mike's amp at the final volume (or external pedal volume) is about 30Vpp. The concertina has no gain and must be fed a signal that is essentially the same amplitude as required for the output tube grids. The EL-84s have a sensitive grid, so just 15Vp needed. Octals would require two to five times this amount, depending on the tubes and the power output. So, for the latter, it is good to have a gain stage between the MV and the concertina, or use a higher value pot for the MV and pedal.

Also, it might be worth adding a disc cap from rail to rail at each stage, say 100nF. With certain circuits, there can be distortion transients with low-level signals that do not occur with high-level signals.

When I first saw the schematic, the input stage reminded me of the Laney Klipp, which uses a similar circuit for its OD section. The way the gain pot is wired is also common in a lot of jfet preamps, like Traynor's BLOC-series from the 1980s.One end provides unity gain while the other has a boosted signal, so there is no zero signal point on this particular control.

Oop! Just had a closer look at the schematic and the bottom of the first gain control goes to ground, so it is a simple level control that does go to zero loudness. Thought it was tied to the cathode... need new glasses....

Have fun