I didn't much care for the sound of my active elliptic filter - great dynamics in the bass for sure but the upper-end colourations were a bit unnatural sounding. So I've shelved tthat one for now and instead I'm playing with a simplified (by which I mean fewer inductors) passive elliptic.

There are two topologies for building elliptics where the zeroes are realized either by shunt series-LC networks or series paralleled-LC networks. The series created zeroes means fewer inductors are called for. In its most basic, unbalanced form there would be just three inductors for a 7th order filter. This filter though is balanced and designed to feed my Nitro desktop amp directly, without any I/V amplifier stage.

Passive filters rock for SQ, no doubt about it but I'm still curious how good sounding an active DAC I/V post filter might be. So I've figured out an almost equivalent FR active version of my 7th order LC elliptic filter. This active elliptic has been designed using LTSpice's FilterCad program giving the pole/zero positions, then the Williams handbook of filter design helped me translate those numbers into a working circuit. Its using what Williams calls the VCVS 2nd order section based on a twin-T network to realize the zeroes.

My first attempt at an active elliptic filter was using gyrators but that proved very hard (practically impossible) to get stable with CFB opamps due to their HF gain peaking. VFB opamps I ruled out at the start for inadequate SQ - its not hard to make gyrators stable with them. Hence this approach which promises to work with CFBs though I'd guess I'll probably need to add series Rs between the stages in practice. Nothing built yet but thought I'd...

I've recently come to the view that the best metric of SQ in DACs is noise modulation, however as yet I'm unclear how best to go about making the measurement. Stereophile tried, and here's what they did - Noise, Modulation, & Digital/Analog Conversion | Stereophile.com

Robert Harley reports no correlation with listening so I deduce from this that the test method isn't looking for the right thing. My own gut feel is we need a multitone test waveform rather than a single sinewave as a sinewave has a crest factor quite unlike music and if a sine provoked the condition we'd see it on the 'N' part of a THD+N vs level test.

Having said that, of course we do indeed see something weird going on - on Weiss Medea's implementation of the ESS Sabre. The THD+N plot versus level shows kinks which most certainly are not increased THD at various levels, so can only be noise floor modulation.

Anyone have any ideas?

For those who missed Frank (fas42)'s link on a thread I started then here's where I'm continuing my minimal oversampling DAC developments for the time being : Digital that sounds like analog

Here's one of the clearest articles I've seen which explains the background and a couple of solutions to DAC roll-off.

Flatten DAC frequency response | EDN

Both the solutions proposed compromise the dynamic range. My own solution, a hybrid analog/digital approach relies on DACs being cheap - I've termed it 'LAID' which stands for 'Longitudinal Array of Inexpensive DACs'.

The prototype is listenable now, but a bit too much background noise/hum pickup to do serious listening. It needs an on-board post filter and amp, which I'm working on now. In the meantime, here's the pics - the DAC itself is built of 5 'dac-sleds' each with a stack of 4 chips. The 'sleds' are then arranged around the central tower holding the resistor ladder. A separate board handles the timing logic and 18 tap delay line.

Fixing the NOS droop has the undesirable side-effect of making the imaging components worse - HF gain can't be increased up to 20kHz and then suddenly taken away above 22kHz

In the spirit of taming the near-ultrasonic emissons of a NOS DAC, I'm currently playing with an MOS design - where 'M' stands for 'minimal'. I've been wondering if the attractiveness of the NOS sound will still be preserved if I go to 2X OS in order to fix up the ultrasonics. No practical analog filter can have a sharp enough band edge so a digital one it does have to be...

A DSP implemented digital filter comes at a price - that of throwing away some bits (I'm still using only 16bit DAC chips - TDA1387) so I'm now exploring using a transversal filter (no DSP) to carry out 2X OS. The LTSpice screen grab shows the architecture I'm playing with - a 19 tap delay line feeding 19 separate DACs. The DACs are shown on the right as current sources and their individual weightings are...

Plenty of audio DIYers have discovered what an excellent modding base the DAC-AH Lite is. I'm a latecomer to the party, but better late than never here's my contribution to the art of DAC-AH-alikes.

Most of the mods I've seen focus on the output stage, some upgrade the PSUs. Here's yet another variant - change the DACs. The TDA1545 is pin compatible with the TDA1543 so I swapped out 8 TDA1543s and replaced them with 8 DAC-stacks made up of 5 TDA1545s each. 32 DACs in parallel produce the normal NOS output, the additional 8 DACs are fed from a two-stage delay line (16 74HC595s) to correct for the zero-order hold. I/V conversion is done passively and differential - 1.3ohms in each polarity. The post I/V stage is a pair of AD605s giving a true differential output. In between the two is a passive filter made up of TDK ferrite beads and NP0 capacitors.

How does it all sound? - well of course one is always biassed when describing one's own offspring, but in a word...

If you have a NOS DAC (like a Metrum) and are curious to hear how it sounds with the zero-order hold corrected, here's a fairly simple way to try out my 3-tap filter using Audacity.

Load up the file you'd like to process and make two duplicates (select original and use Ctrl-D) - we'll call them A and B. We need to apply delay and gain to A and B which we do with Audacity's time-shift button (in the group of 6 tool buttons to the right of the transport buttons) and the 'Effects - Amplify' feature. A is shifted one sample to the right, and B, two samples right. I've shown how this looks with a mono chirp signal in the first image.

The gain is input into Audacity in dB so 0.15 becomes -16.5dB and 0.026 is -31.7dB. These are the gains for A and B, respectively. A needs to be inverted too - use 'Effects - Invert'. The second screen grab shows how it looks after these operations (I hid the device toolbar to give more space for waveforms).

Having processed...

Here's a ten discrete tone test waveform played back into my Sony PCM-M10 then FFT'd in Audacity with 512point FFT.

Audacity reports lower freq tones at -18.6dB and the highest (17.3kHz) at -18.9dB - a droop of 0.3dB. This might be in part my passive (LC) reconstruction filter which I have yet to characterize separately. So it appears to work