Designing the best bang-for-the-buck headphone amp

Headphone amps aren't any different from speaker amps in principle - what you hear (apart from a bigger version of the input signal) is the power supply's noise coupled through the inadequate PSRR of the electronics.

SE classA operation is a way to minimize the generation of power supply noise by arranging the current flow to be constant to a first order so that any remaining ripple on the supply is the result of the finite output impedance of the follower's loading current source and those of the driving stages. But how significant are these 2nd order effects? This design is an attempt to find out - by reducing them as far as practicable.

The idea is to run SE classA at a much higher voltage than is needed to drive the 'phones (balanced, with 80V supplies giving 144V peak-peak) then step down the output voltage with a custom-wound output transformer. This has the effect of increasing the PSRR of the amp's output stage. I'm not worried overmuch about the PSRR of the earlier stages because not handling any significant current their power supplies don't need to be such low impedance. Running at high voltage also it means running at lower output currents so hopefully the semiconductors will be more linear too. The downside is more gain is needed because the trafo has negative gain (in dB terms). Another upside of running at higher voltage which is interesting to me is that the capacitance required for a given power supply ripple gets smaller (in uF and also in physical size and cost) because the energy storage of a cap is a function of the square of its working voltage.

Having the output transformer brings another major benefit - with multiple secondaries its possible to optimize the drive levels for a multitude of different impedance phones without wasting too much classA bias current. Without the trafo, the bias levels would need to be either variable according to the connected load or set to drive the lowest impedance with concomitant waste when driving high impedances. I've wound one of the output transformers and the second is underway. These transformers are seriously cheap in terms of materials (on Taobao the PQ32/30 core sets are about $0.50) - the major investment is in terms of the time taken to wind (ignoring the cost of a little wire).

With the first trafo I pretty early on got tired of counting the windings - I always lose track anyway - and this winding is the longest I've so far done, with a primary resistance of 1.3k. The wire is 0.07mm diameter which clocks in about 3.4ohms per metre length, so there's almost 400m of wire there. The calculations I made showed I'd need about 4900turns. Due to estimating the turns by measuring the diameter of the enlarging coil I actually ended up with about 5200, calculated by applying a known AC voltage and seeing what came out of a test secondary winding of 10 turns attached to my millivoltmeter.

The secondaries are much easier to handle, with the proviso that all four of them have to have precisely the same number of turns (141) so they can be wired in parallel to drive the lowest impedance phones with a respectable source impedance. They're made from 0.23mm diameter wire. Wired in series they'll put out about 100mW peak into a 600ohm load when the primary's driven to 72V.

So what bias current is necessary for the OPS? Its a function of the worst-case loading and also the magnetising inductamce of the trafo's primary which increases the current draw. On the LCR meter the primary inductance is 140H however this is somewhat level dependent, the ferrite material not being particularly linear. So I'll make some measurements when I have a working channel (i.e. two amps in anti-phase). I'm not yet sure how to do the phase splitting - perhaps an input trafo?

For a power supply I've re-wound the secondary of a 5VA EI transformer to hopefully give me around 90V unregulated which I'll use an LC then a series pass transistor on. Secondary resistance is 50ohms, primary 675ohms. I suppose if I had enough patience to wind a very large choke, a choke input supply would work - trouble is it would need a higher secondary voltage. That might be an experiment for a later time.

I've uploaded the schematic - there eventually will be opamps where the two voltage sources are, arranged to deliver anti-phase signals. The 1.75V sources (with 68R in series) are in practice red LEDs. The output is strung between the two phases - shown as a 52k resistor and the shunt inductance and capacitance of the transformer (140H, 65pF).

Of note are the 'Hawksford' style cascode for the top side current source, except I've tapped off its feed point to avoid instability. Similarly the darlington buffer is bootstrapped and its stability is ensured by the RC network to neg_rail.

One simulation result of interest. To see how much power supply noise is generated I introduced a 100R series resistor in the rail. But turns out LTSpice is better at showing the current through that resistor than the voltage induced on the rail. So I've added a plot of that current - it varies between 7.243mA and 7.24294mA with a full-scale 40Hz signal applied. So just 60nA peak-peak.

Pic of first prototype added - not yet powered. The TO126 transistors are 2SA1381Ds and the TO99 cans are LM144Hs.

Update - I've been listening to my mobile phone through it, into the AKG240s and Superlux HD668s. One of the first things I noticed, trying the Superlux first (I figured they're cheaper so if it proved unstable, less to blow up) was they reminded me of listening to Stax's, many years ago. The way the soundfield is so disconnected from the drivers that produce it was the main resemblance. At first unimpressive - which to me is a very good sign. There's so much ambience retrieval its difficult to tear myself away from listening. This amp sounds less like an amp of anything I've built to-date - quite hard to distinguish between sounds coming from outside and those on the recording as the soundfield is so 3D.

Attached Thumbnails