Author: Dmitry Nizhegorodov (email@example.com). My other projects and articles
3. Basic schematic
4. Basic findings
5. More advanced investigation
6. 120 Hz Hum elimination: DC on filaments
7. Diode Bridge and CRC filter
9. Regulated supply: voltage or current or both?
10. Bridge into regulated current supply.
11. Doubler to regulated current
12. Diode model
My "Guinea Pig" was Welborne Labs Moondog 2A3 amplifier. Note that I don't claim that Moondog hums more than other comparable amplifiers, nor I claim it is better in this or any other respect. Moondog is an excellently sounding SET. It just happens that a typical SET with DHT filaments hums.
I consider Moondog having a solid, somewhat "conservative" design. Moondog is an ideal bench for studying the particular source of hum we describe here - hum of other nature in Moondog is negligible. It has DC-fed amplification and driving stages and a beefy choke-based CLCRC B+ power supply (PS), producing very low PS hum.
Moondog provides at least 3.5 WRMS of output with fairly low harmonics and exhibits classical SET spectrum: dominating 2nd harmonic, rapidly diminishing higher products.
This hum is not "buzz". It is a low-harmonic tone.
Shorting the inputs did not eliminate or change the hum.
Changing grounding on the AC line (120V power) sockets did not reduce the hum.
Balancing of the filaments with balancing pots does not neutralize that specific hum although it eliminated a differently-sounding hum sound, audibly lower in frequency, Both monoblocks produced the same level of hum.
I opened the amp. The layout appeared to be reasonable, and wiring was fine, as a friend of mine who owns these Moondogs assembled them very accurately, following the wiring instructions precisely. All AC wires were properly twisted and positioned.
Removal of the first 2 triodes did not reduce the hum. I grounded the grid of the output tube, and that did not reduce the hum as well. Shaking the wires and passive components did not change much (although I found that the balancing pot of Moondog is super-sensitive to very small rotations and even vibration).
Next, I connected a scope to the output, watching the level of hum and replaced the speaker with a load resistor, observing the same level of hum. Then I connected a CR/R DC-blocking and dividing circuit to the plate, which allowed me to see signal more clearly:
On the scope the hum measured at ~ 30mVRMS on the plate of 2a3 (taking into account correction).
The first picture is with unbalanced pot, the second one with maximum of balancing possible. Note: this equilibrium is very unstable on Moondog. Perhaps, the pot needs be limited in range.
The hum from unbalancing has frequency 120 Hz (2 * AC). The hum has frequency 120 Hz (2 * AC).
Thus we have about 30 millivolts of the 120 Hz signal. This is in the US, where the mains are 60 Hz. Note that the hum-balancing resistor eliminates the 60 Hz signal, whereas this 120 Hz residual seems constantly present. Indicative of power supply ripple, which is 120 HZ? That was my first thought. I introduced a switch cutting off B+ from output tube, the rest of schematic running as is. No effect on hum. I introduced another switch, cutting off both wires going from filaments of filament supply of 2a3. Now I could see and hear a difference: removing the AC from the tube's filaments INSTANTANEOUSLY killed the hum.
RESULT 1: 120 HZ hum is coming from the filaments, not ground, in, or B+.
In part Moondog makes it easier to investigate this simply because its PS is pretty very well filtered: 4-6mV RMS of hum on the B+rail.
The schematic on the right is the stock powers supply for Moondog, sans the RCRC chain for the input stages. This schematic is for SPICE simulation. Rampls is the resistance of the amplifier. Cdcblock and Rprobe are needed to remove the B+ voltage from the probe.
This plot show the residual B+ hum of the PS (Vdcblock). The second curve - V(V4)/20k is the AC signal before the rectifiers, scaled down by 20000. Note that the residual hum is very much sine-like. Note the phase shift: approximately 65 degrees.
When the filaments are disconnected, the B+hum is not audible; indeed it is only 200..400uV on the speaker terminals.
Steve Bench suggests  that 120 filaments hum is the result of non-infinitely heavy filaments feed with 60Hz AC, thus heating up and cooling down with 2x of power frequency. While this effect clearly is a part of the equation (imagine a very low filament frequency, say .1Hz, and very tiny filament, we then can see in real time how it heats up and cools down) I found that for power tubes such as 2a3 or 300b in normal steady class-A SE configuration (below-power-dissipation equilibrium current, non-starved filaments etc) have negligible thermal hum. Practically all hum of 2a3s or 300b in audio class A circuits is due to tube non-linearity. Among several tubes the one that is most liner under operation conditions is the one that will hum less and not the one that has heavier filaments regardless of linearity. See  for more details.
Because hum seems a property of tubes, not circuit, we could try reduce the hum by tube selection. However, obtaining very large quantities of 2a3 is problematic and expecting exemplars with exceptionally low distortion is unrealistic.
Another direction is to assume that typical tubes are used find engineering solutions. There are 2 directions to go: (1) use hum elimination techniques (2) use hum-cancellation techniques.
With the first direction, we remove the source of the hum - the low-frequency AC signal. The most obvious solution here is to switch to DC filament supply. There is another, more complex solution - move the filament AC frequency well above the audio range, say, run it at 100 KHz.
In the second category are various hum-injection methods similar to those proposed by John Broskie and Steve Bench. In the next sections I evaluate my results in these areas.
There is a widespread belief among triode enthusiasts that DHTs are better sounding when compared to indirectly heater triodes (IHT). More, among those purists who value DHTs, filaments fed with AC are often preferred over DC; those are considered troublesome and/or sounding inferior. It is often claimed that DC filaments increase harshness while reducing or wailing other desirable properties such as soundstage and image quality and dynamics. Engineers, who often dismiss sonic superipority of AC filaments, still often admit that DC on filaments reduces tube life and performance over long time. Also, engineers clearly see lots of added complexity of DC filament supply: part number, space occupied, heat dissipated. design and implementation.
Is switching noise a real problem? Yes. Low-voltage, high current supply can generate huge current spikes. Often current spikes in a diode bridge for a 2.5A supply go above above 20A.
Use of low-noise (super-fast or Schottky) diodes improves the picture but bad audio quality of a capacitor still remains a widely claimed problem spot. In this section we estimate how a "generator" placed across filaments propagates signal to the plate load; this hints that such claims may be questioned. We'll come back to this issue laterm and for now let's assume large capacitors arer best be screened from DHT.
One promising direction is CLC or CLCL filtersing. A choke for such application must be high-current, low resistant and sufficiently inductant; price/size/weight of a 3+A, 20+ mH choke which is less than .2 ohm may be unacceptable for many designs. Indeed, Moondog's chassis has no place to go fosr such choke - too crowded on the top, no enough room under the hood. Still, below is SPICE simulation of such design with a 20mH choke.
What diodes to select? One example is STPS20L15D, a 20A, 15V diode (digikey part# STPS20L15D-ND). It drops less than .25V under 10A and when junction temperature is above 75C.
Following this route, can a doubler-plus-choke DC supply be made? The answer is yes, here we use the same diodes as above.
The choke isolates the tube from the diodes and removes hash, and helps to reduces ripple. As larger the chokes a better the job - a 1mH one is equivalent to ~1ohm at 120hz. The following is a working design decision, using 2 1mH chokes.
A 1mH choke that can withstand 3A needs a lot of iron; sich chokes are heavy, big and pricey. ElectraPrint sells quality filament chokes for 2a3 or 300B. Yet there is an alternative. A small-core common mode choke can be tried instead, since the current float is opposite directions. One example is J.W.Miller 8107 choke (digikey # M9842-ND, Mouser # 542-8107), a 2mH 6.6A 0.022 DCR choke. Another attractive candidates are J.W.Miller 7123 (11mH 5A 0.072DCR) or 7122 (30mH 4A 0.13DCR).
this filament supply results in < 200mV ripple and with the plate of 2A3 develops less than 1 mV of 120hz hum!
Note use of a choke is not obligatory. With resistors R1 and R2 set to 0.15 the signal on the large caps is ~3V and pretty nicely shaped; by adjusting the resistors this can become 2.5V. The idea can work even with smaller caps:
residual hum is ~ 0.6VRMS. If the tube is wired with a balancing pot across its filaments, the hum will be cancelled out and virtually undetectable.
The following is a working choke-free design decision.
with the caps and resistors specified, this filament supply results in 2mV of 120hz hum on the plate of 2A3 - not as good as with choke, but still very good.
N.B. Accurate plate hum simulations were made possible with my composite models for directly heated triodes, see . The 4-pin 2a3 PSPICE model shown above is described there.
Use of current-limiting resistor hints that a current regulator, not voltage regulator maybe be a better solution. Indeed, current sources for filaments, just like current source for plate loads or for cathodes have reputation to preserve delicate qualities of triodes. And current source is something that will provide the most gentle warm-up regime for the filament, much extending tube life beyond what AC-fed filaments can provide.
A 2-phase regulator can also be conceived, with voltage regulator followed by a current regulator. This approach may provide the best stability. However, sequential V regulator and a current regulator eat lots of voltage and the transformer winding needs to supply that extra voltage. Also, all the voltage will dissipate as heat. Here are some estimates. Suppose, a Schottky diode bridge is used, each diode's drop is .5V. A voltage regulator eats around 1.1V (there are some that take less but ~1V is more typical for LM or LT regulators). Then, the second regulator eats the same. Then comes the drop on the voltage reference resistor. When a voltage regulator is used as a current regulator, it regulates voltage on a resistor that is in series with the load. A typical output voltage of an adjustable regulator is 1.25V and that is how much will be wasted on the reference voltage resistor. Together, we get 0.5 + 0.5 + 1.1 + 1.1 + 1.25 = 4.45. More, to avoid loss of regulation, we need to provide some padding, and an extra 2V for the first regulator and 1V for the second is a must. Thus we have 6.45V of waste. This is a lot of voltage! This is how much more the transformer needs to provide on top of 2.5V AC. This is a lot of heat to waste, too: 6.45V * 2.5V = 16.25. It is as much heat as produced by fully-loaded 6BQ5. You can not hide that heat under the chassis, at least not for Moondog...
If the voltage regulator is skipped, the "efficiency" increases but not by much.
LT1085 is a 3A adjustable voltage regulator, minimal voltage out is 1.25V. In this circuit, it dissipates around 7W and needs a good heat sink. A 10000uF capacitor is sufficient but a larger value can be used to smooth the incoming ripple. However, value too big may result in excess current peaks. The peak current impulse into 10000uF is 12A.
The voltage reference resistor needs be at least 5W. Its value is chosen to provide ~ 2.3 - 2.4 V into the filaments, which prolongs tube life. The value may need be adjusted for specific tube and regulator. The 1U capacitor is added for stability.
On this picture, you can see the LT1085 regulator and a current-in-rush-limiting resistor mounted on a heatsink.
Two white ceramic Dale 10W 1 ohm resistors form the .5R value, and the twisted wire next to them form the additional ~ 0.02R value needed to limit the filament voltage to 2.4V.
The bridge is a PB605 (6A average, 150A impulse, 1V voltage drop per element). The sink gets pretty hot despite being glue directly to the sink.
The transformer is Tamura 20W with 5V secondary; the 7.5VAC shown is formed as 5V + 2.5V. To reduce heat dissipation by LT1085, 5V + 1.25V can be tried, yet in this configuration LT1085 is on the verge of losing stability of regulation if the wall AC drops below 110V (of course, this is with PB605 used as rectifier; a Schottky diode bridge will free up extra 1-1.2V.
This is just a prototype but as you can see, the space under the hood becomes crowded even without the sink which in the complete version must go outside. The transformer gets quite hot and it would better go outside, too.
No balancing potentiometer needs be used with this filament supply; if the amp uses fixed bias for 2a3, the ground of this schematic goes to the ground filament pin. If the amp uses a cathode autobiase, as Moondog does, then the "ground" of this supply must eb left floating, connected to one of two filament pins.
LT1085 dissipates around 7W and needs a good heat sink. The diodes are assumed to have Vf <0.5V. An extra resistor of ~ .1 ohm can be used in series with the regulator to unload it a little bit. However, it will not be the case if the diodes drop more than 0.5V. The voltage reference resistor needs be at least 5W. Its value is chosen to provide ~ 2.3 - 2.4 V into the filaments, which prolongs tube life. The value may need be adjusted for specific tube and regulator. The 1U capacitor is added for stability.
This plot show the filament voltage, the voltage on the regulator output and the voltage on its input. The value of the capacitors must be quite high (10000 will not work).
if extra dropping resistors can be used, consider adding them before the capacitors. This will limit the current peak through the diodes. Without such dropping resistors the inrush current may exceed 20A.
.model D_Schottky_STPS20L15D D(Is=168.1E-11 N=.2 VJ=0.25 RS=.01 )
This represents STPS20L15D at ~80-90 C in region of 1-20a pretty closely, entirely suitable for our needs.
Why it is a pretty good fit is evident from the following. On the left is a clip from STPS20L15D's data-sheet; on the right is spice simulation of D(Is=168.1E-11 N=.2 VJ=0.25 RS=.01 )
data was obtained and tuned in "real transient mode" using this idea:
if you need to develop your own model, first set RS to 0 and adjust low-current region and "slope" of the curve at low currents with IS and N; the plot is a line I is log and U is linear; then add small value for RS and see the plot "bending" (I scale is log, U scale is linear). You'll have to do it iteratively many times until the right shape is found.
Gizmo on filament supply
 composite-triode-models.htm Composite Triode Models
Author: Dmitry Nizhegorodov (firstname.lastname@example.org). My other projects and articles