In a side-by-side comparison between Kubelik and Deca DAC, the latter wins out on dynamics, especially in the bass. I was very curious to find out what was going on to reduce the dynamics in Kubelik.

First up I figured the filter might be the primary reason - Kubelik has a very simple, 3rd order (CLC) filter whereas Deca has a complex 7th order one. So I built variants of Kubelik with higher order filters, firstly 5th order then 7th order. These needed piggy-back PCBs on which all the extra inductors were mounted. Not a pretty sight!

Results were that dynamics did improve but even the 7th order version of Kubelik didn't reach the levels of Deca DAC. So then I wondered if it was down to the opamp I'm using (OPA2209). I bypassed that, feeding the SE output stage directly from the I/V resistor - hence 'passive I/V' rather than active, just for the hell of it. Wow, a very large improvement - much more engaging sound! The level of improvement was very surprising.

So was the lack of dynamics the opamp's fault? I decided it needed a 2nd chance, so I brought it back in, this time as a buffer rather than in inverting ('transimpedance') mode as is traditionally used for I/V. The engagement factor stayed. Hmmm.

In simulation it turns out that the output impedance of the filter interacts with the opamp to increase the noise gain at certain frequencies with a traditional I/V stage. With passive I/V, the opposite happens - the noise gets reduced at those same frequencies. Seems like this is a plausible reason for the improvement in dynamics going from transimpedance to passive I/V. So then I wondered how much control I had over the frequencies where the noise was being reduced. Naturally, it depends on the filter and I found a 5th order filter had potential to reduce the noise close to the most sensitive range of our hearing (3 - 5kHz roughly). The Fletcher-Munson curves for our hearing thresholds vs frequency have a noticeable dip in that region. However, for some reason I also turned up ITU-R468 which is more relevant for noise (F-M is based on a single tone). In that standard, the peak of the weighting curve is a bit higher in freq, just over 6kHz.

I wondered if I could design a filter which gave the lowest noise in the vicinity of 6kHz. From fiddling with the Chebyshev online filter designer (https://rf-tools.com/lc-filter/) the dip in the noise turns out to be a function of the ripple of the filter, not just the cut-off frequency. Going higher ripple allowed me to place a dip in the noise gain very close to 6.3kHz.