New Headphone Amplifier Design

Thanks, Hugh. 500th reply. Nice.

Boards are in production at this moment and I'm taking next week off so I will have free time. If things go well, I will get this built shortly.

That would be pretty good, less than three weeks from concept to completion. One day -- it's like a miracle -- it will appear.

I'm listening to Coleman Hawkins on the A3 right now. Sounds nice. Thank you, Jeff.

I am working on something similar with a MOSFET output, but not able to get the loopgain profile like yours. Can you take a look and comment..Please.


  • Tube Buffer_HPA_MOSFET_Final.asc
    15.5 KB · Views: 10
Here is an updated schematic. I will ask your advice here. The issue is with input DC offset. With nothing connected to the input, excepting the 100K resistor to ground, both channels pull the input to around -700mV and there is a corresponding 2.1V DC offset at the output. Shorting the input reduces the offset to a level that the servo can correct (though it's working hard). The trimmer potentiometer doesn't have enough authority to zero it out.

Evidently, the input stage is drawing about 7uA unbalanced bias current. I wasn't expecting the difference to be that great. There's DC flowing in the volume controls and you can hear some noise when changing the settings. Otherwise, the circuit seems to be working as expected.

I need to think about how to fix this. It's not rocket science, but I want to understand better what's going on.

That would definitely not be a good idea, removing all the DC feedback.

The problem is DC offset at the input. The only way for R15, the 100K input bias resistor, to develop a voltage across it is for a net current to flow in/out of the input transistor bases. Q10 and Q12 emitter currents are close, about 5mA, based on the voltages across the two 22 Ohm resistors. Vce is about 15V for both transistors.

I plotted collector curves for a random BC550/560 pair. Granted, this doesn't go all the way up to 15V, but extrapolating, it's hard to see where we come up with a 7uA base current difference. Still, it's not inconceivable. The current has to be coming from somewhere.

I guess I will need to capacitor couple the inputs and inject an offset current there, and get rid of the existing offset pot.



2010-11-18 9:12 am
You're wrong.The output is the one forcing the emitters of the diamond's feedback trz through dc feedback . Simple mathematics.Nakamichi used cap coupling the feedback in some of their most expensive and cherished products.Tons of nichicon muse everywhere...the reason for switching from ac coupled design to dc coupled designs was mostly the general low quality of electrolitics in the 70's .2 decades later dc coupled designs improved a lot same as the electrolitics did. With today's electrolitics i'd scrap dc coupling everywhere if possible.I don't give a damn on those telling they need -3db at 5 hz to have a clean base...That's simply delusional . 3 years ago I measured a 10 k $ Chord that was reveered by some high end headphones manufacturers and I was surprised to see it actually worked well only from 30 hz up...Apparently it helps the Harman curves...
Last edited:
Intending no disrespect to dreamth, I don't think his analysis explains the problem. A 2V offset at the amplifier output corresponds to a 20uA net current from the driver stage (into the 100K load resistor, R18). The current gain through the input buffer is about 250,000 and this is more than high enough that the resulting input current should be very low. I understand the limits of SPICE. Still, I ran a simulation of just the buffer with a 5mA current source on the output and the resulting input offset was nowhere near what I'm seeing. In practice, the current flowing in/out of the inverting input is very, very small -- because of the high open-loop gain.

The only answer I can see is that the DC current gain/offset between Q10 and Q12 is large enough to give this 7uA error current at the chosen operating point. If the input transistors were perfectly matched, the base bias current flowing out of the top transistor would exactly equal the current flowing into the bottom transistor and the total offset current would be zero. This assumes equal emitter current, which is very nearly the case in my amp.

I've been working with JFET inputs for the past year and a half and didn't really think about the difference using BJTs. Evidently, this was an oversight on my part.

I use a 100K bias resistor (R15) because my volume pots are 20K and it doesn't make sense to hang, say, a 10K resistor from the wiper to ground.

R36 is 100K to try to isolate the op-amp, but it could be made much smaller. However, it won't change the situation at the input; the same current will flow from the op-amp, just at a lower voltage. In any case, I've been testing with the servo jumper removed, so the op-amp is not a factor.

FWIW, the amplifier sounds very good. It's hard to tell, and there's the whole maybe-it's-real-maybe-it's-not break-in thing. But it kind of sounds like the A3, but with a more incisive top end.

I'm going to have to do another iteration on this, but I guess the first thing is to hack this board to try to make it work better before I go off ordering more PCBs.
The current mirrors aren't the problem. Increasing RV1 to 500 Ohms would screw everything up royally.

I did some serious debugging last night. I found that the first stage bias currents were about 5.5mA as expected, but the second stage mirrors were only providing about 1mA. Oddly, the second stage emitter currents were about 2.2mA. I have absolutely no idea where the extra current was coming from. Spent a lot of time trying to figure it out, too. I decided to change the emitter resistors (R10/11/20/21) all around to 100 Ohms. I also changed the servo summing resistors to 22K. With this setup, second stage current increased to about 4.2mA and the bias servos aren't pinning to the rail. I was able to trim one channel to zero offset with the servo jumpers disconnected, but could only get the other channel down to about 100mV. However, it settles to a low value once I connect the servo.

There is still some variation in offset when I adjust the volume controls, which takes time for the servo to correct. And there is hum with the volume controls at mid-range. And scratchiness when I move the controls.

I'm listening to it now and it's very, very nice. It's warm and transparent, with good bass and extended natural highs. Really inviting sound without the slight muddiness of the A3. But I'm not happy with this input stage, so need to think about it.

Increasing the bias resistor values reduces loop gain about 4dB.

I'm impressed by how good this sounds. It's hit a sweet spot between being detailed and revealing without being fatiguing. So, good potential here, but needs more development.
Last edited:


2010-11-18 9:12 am
The current mirrors aren't the problem. Increasing RV1 to 500 Ohms would screw everything up royally.

I did some serious debugging last night. I found that the first stage bias currents were about 5.5mA as expected, but the second stage mirrors were only providing about 1mA. Oddly, the second stage emitter currents were about 2.2mA. I have absolutely no idea where the extra current was coming from. Spent a lot of time trying to figure it out, too. I decided to change the emitter resistors (R10/11/20/21) all around to 100 Ohms.
Haha...I asked about modifying RV 1 (which is in series with R10 and R11 ) modified to 250...500 ohms and you basically made r10+r11 =200 ohms... then mention only the other max value of 500 ohms as unsuitable...Have you EVER admitted to ANYONE being wrong about ANYTHING in your life?
Increasing RV1 to 500 Ohms while leaving the other resistors at 22 Ohms would screw things up.

I'm not in the mood for an argument. I appreciate your suggestions, but you don't seem to understand the problem, so what can I say?

Edit: And for what it's worth, evidently, I was wrong about my initial design, which is why it doesn't work right.


2010-11-18 9:12 am
I don't have your design in my sim to check it right away...I just assumed those resistors need to be larger in value to allow some play range and force the input balance to zero and you had a totally inadequate value for RV1 which was striking even without seeing it in a simulator.I suggested the usual 250..500 ohms trim values as they are standardized for multiturn trim pots.It doesn't matter if you screw up things in one stage with a 500 ohms trim pot ...any electronics design allow for a large variation of different parameters so that screwing one stage can be fixed by screwing one more stage.With this I came back to my previous and older assertions that your designs look more optimized for an 4...8 ohm 20 watts load than for headphones.With 250 000 current gains you can drive just about anything including F22 raptor jet engines man!
I don't understand why you say the 20 Ohm trimmer is inadequate. With the values originally specified, this lets you adjust the total emitter resistances to be 12 and 32 Ohms (or vice-versa) at the limit. This will radically unbalance the second stage, more than enough to trim out any conceivable output offset.

SPICE says it should work very well, but SPICE isn't reality. I don't fully understand what was going on with the 22 Ohm resistors, but it seems like the Vbe mismatch was making problems and I need more voltage drop across the bias resistors to get the circuit to behave. I really, really want to know where my extra 1.2mA was coming from because it seems like I was violating Kirchoff's law and that bothers me.

Ideally, the input buffer should have zero output impedance. Increasing the emitter resistors raises the impedance and introduces an extra term into the CFA gain equation. This isn't really a problem in practice, but it's less "pure." Bonsai uses 150 Ohms in his designs and now I understand why.

I will probably remove RV1 from its current position and instead wire up a variable voltage divider between the supply rails and connect that to the bottom of R15 so I can null out the input offset. This is what the RJM Sapphire does, for instance.


The servo can take care of any other DC offsets in the circuit, maybe.

Another option is to change the input devices to complementary JFETs. This would eliminate the input bias current.

I need to study and think about this some more. I've been listening to the amp for a couple of hours and I'm really impressed with how it plays. I have a chassis ready to go, but I'm not going to build this up until I sort out the kinks.


I know your position on powerful amplifiers. I probably said something like this before: It's a big world and there's plenty of room for people to experiment with what interests them.