Tweaking electronics sound quality in active speakers

My main listening system has been based on Fayou XPE active speakers for over six months now. I bought my first pair of Fayou speakers a year earlier than that, the lower-spec'd B2 models. The B2s are now my main PC speakers, connected to an Asus Essence ST. After getting the B2s, it was only a matter of hours before I had them open, poring over the implementation details. I noticed that the grounding to tweeter and bass/mid was shared so sliced into the board to give each unit its own dedicated ground path. I also tweaked the high current paths to the main PSU caps to ensure no impedances in common with the audio signal grounds. After these mods, the sound was highly enjoyable.

When I first heard the XPEs I considered them not to be quite as good sounding as my (tweaked) B2s, but given their superior amplification and bass/mid driver, figured they were ripe for hot-rodding. Their construction also suits tweaking much better as all the electronics is easily accessible on the inside of the backplate heatsink. A review (in Chinese but Google is your friend) with some great pics of the innards is here. The price is a very reasonable 1250RMB for the pair, that's under $185 at today's exchange rate.

The XPEs are also more attractive because they use twin LM3886s, a chip I've been familiar with from around the time it became available in the RS catalog in the mid-1990s. Having prototyped active speakers using this chip, I'm highly enthusiastic about its capabilities in the right implementation.

Reverse-engineering the design

There are three PCBs - the first being one I've so far neglected - the mains filtering functions. Second board is the input buffer and volume control, using a single NE5532. The third and most complex is the amp, crossover and power supply board, that's the one on the rear heatsink in the photos.

Beginning with the input buffer, it turns out on closer examination that half of the NE5532 is unused, so a 5534 could have been employed. Guess this pcb is a common one so gets re-used in other designs where there's a stereo input. The XPE is odd in this respect, its totally a mono design which means two (unganged) volume controls, rather a user-interface nightmare. The input buffer is a 2nd order high pass, Sallen-Key topology, I put the component values into LTSpice and it plotted the frequency response. A bit too high-Q for my taste, so out came the 1dB bass peak by increasing the size of the feedback resistance.

The input's RF filtering was a bit hit-and-miss, a 680pF ceramic after the volume control - this makes the roll-off volume dependent, not quite what we want for the best RF rejection. So I moved the input filter right to the phono socket, adding some series resistance (can't recall exactly, 1k?) and 2 * 100pF 1206 to ground. Output of the high pass went through a 4u7 electrolytic but no series resistance, so I exchanged the cap for a 47R to ensure stability driving the screened cable to the main amp board.

Last stop on this small pcb is the decoupling arrangement - two 100uFs from either rail to ground. I'm not in favour of this arrangement as it dumps power supply hash onto the signal ground, and with only 4 wires in the interconnecting cable there's no chance of adding a power ground wire to the signal ground one. So out came the two caps to be replaced by a single one (twice the voltage rating) between the +12V and -12V rails.

The high pass filter caps looked to be as cheap as possible, but I'm not yet convinced that capacitors make an important difference to the sound, so I left them in pending listening tests.

Main board- power supplies

A centre-tapped secondary on a 150V toroid feeds a bridge and then into 2 * 10,000uF giving an off-load voltage of +/- 31.5V. These rails are for the LM3886s, the rest of the active circuitry is fed from 7812/7912. I considered initially replacing these with LM317/337 to get improved ripple rejection - the 317s allow decoupling the reference with a cap giving up to 20dB improvement in ripple. But that's a bit of a dog to implement on this board so I took a closer look at the circuit.

One of the problems with 3pin regs is not so much their poor ripple rejection in the audio band, but how fast this degrades above 20kHz. For good sound, we need to keep the supply rails clean into RF frequencies and 3pin regs have very poor rejection there. So they do need input filtering and this hasn't been included in this design. Fortunately there's a lot of voltage headroom available, so series resistors are a cheap and simple solution. In go 2 * 220R in the supply rails between the main caps and the regulators' local decouplers. The second significant problem with 3pin regs when driving low current loads such as opamps is output impedance, particularly at higher frequencies. The polyester output decouplers aren't too much of a help, so I exchanged them for 220uF electrolytics (about the biggest which would fit in the available space).

Grounding

Surprising as it may sound, in my experience fixing grounding has had the greatest improvement on sound quality. I'm not totally sure I can hear differences between opamps, but comparing a poorly grounded design to an optimally grounded one to me is like night and day. Sorting out the grounding on somebody else's design though is not for the faint hearted and in this case its made considerably worse by the board design. They're using a black solder mask which makes it almost impossible to follow the tracks. Perhaps this is deliberate to deter the 'shanzai' (cloners) which are of course rife here. The board is double sided so tracks disappear beneath components.

Beginning at the power source, the grounding around the reservoir caps is obviously not star-shaped. That's easy to spot since they kindly run extra thick solder along the high current tracks, presumably being blissfully unaware of the considerably lower conductance of solder when compared to copper. There are tracks on the top side too to catch out the unwary, some of which have not been masked so the first step is to whip out those big electrolytics to see what's going on underneath. They're glued down so that's just another hurdle to surmount.

I decided to institute a star ground at the point where the transformer's centre tap hits the pcb. The two caps are the biggest current consumers so there must be no other connections direct to the cap grounds themselves. Which meant cutting a few tracks and then re-linking these grounds back to the star with some wire I stripped from a Cat6 network cable (23AWG solid core). The ground is a bit tricky to implement - solder down the higher current carrying wires first (this means power grounds) then move on to the signal grounds. It develops a 3D appearance but that's probably optimal. Don't keep the iron on too long when adding a new wire or the old ones will fall off!

In order then, the two drive unit returns are first, then the power ground return (count as power due to the presence of decouplers) from the 7812/7912 regulators. Then the input board ground wire and finally signal grounds to the LM3886s and NE5532 active filter. I think I ended up with about six wires into the star. I also omitted to cut one track direct to a reservoir cap ground - on powering up there was a loud hum from the bass/mid. So if you get that, you'll know you missed disconnecting a dirty ground. Oops, I forgot to mention there's a seventh entry to the star, but I cheated ever so slightly and returned this to the top side - its the grounding to the heatsink. This originally went via a res cap ground on a very wide top side track so the craft knife is doing sterling duty in isolating this connection.

Component changes

Having been forced to remove the res caps to following the grounding, I decided these were perhaps a bit on the wimpy side for a speaker with audiophile pretentions, so I junked them. I guess that power supply caps do make a difference to the sound, but I've not yet done any listening tests. So I dumped them on technical grounds - they're a bit marginal being only 35V when I measured 31.5V with my multimeter which presumably responds to the average, not the peak. They also felt rather light, so I've splashed out on some Nippon Chemi 'Audio' caps - considerably more expensive than the standard ones at $4.50 a pop. But they're reassuringly large and heavy. These caps are the only major component spend and take the total cost to just over $200.

Decoupling

It turns out that my views on decoupling are a little bit controversial. See for example Best sounding dual opamps. So perhaps not everyone's going to agree with this, but I find that decoupling does make a difference to sound quality and in general the best sound is when there's no corruption of the local signal ground by noise from the power supplies.

This amp has the LM3886s decoupled via polyesters (they're bright orange in the pics) from either rail to the local signal ground. In other words, they share a ground with the 3886's non-inverting gain setting resistor from the -ve input pin. This is double plus ungood because the 3886s don't have regulated supplies - there's a clear path for RF hash all the way back through the toroidal transformer to your home's mains wiring. If you don't want all that electronic rubbish applied to the input of your poweramp chip there are two ways to sort it - either replace the two caps with a single one between the rails, or provide the caps with their own clean ground all the way back to the star point. I decided to go for the first of these options, fitting a 0.47uF polyester as I have a few to hand. Not having a local cap to ground seems not to bother the 3886, there are no audible signs of instability or changes to the offset (did not examine it with the scope).
...