Is there more to Audio Measurements?

[scottjoplin]

Yes, that rings a bell

 

[mmerrill99]

Did you pay attention to all of Martin's presentation. Walking voltage slowly through A/D codes is hardly my idea of "the same way as music".

Sure did & his explanation for how the state variables reach stability when dealing with complex music signals is the underlying issue. It's better explained in Mallinson & Dustin Foreman's paper
"Technical Details of the Sabre Audio DAC"

The consequence of this is that in a quiet passage of
music the state variables of the modulator are all operating
within a certain “state space” and the quantization noise
shaping is described by the noise characteristics in this
“volume” of the space. After a large music transient has
passed, the output traces its dynamic response back to the
quiescent operating point as we expect, but every state
variable is also following its transient response back to its
quiescent point. * During this multi-dimensional excursion
back to the lower signal level the operating point traverses
different volumes of the space, each of which has its own
noise characteristic. Hence a very perceptive listener can hear
something “anomalous” related to the transient response.

* The operating point in the machine state space is returning to the quiescent position – all the dimensions of the state are changing.

He continues..

We
have designed our HyperStream modulators to exhibit strongly
damped responses in all state variables. This means that a low
level signal processed just prior to a transient and just after a
transient, is processed in the same corner of the state space of
the modulator and hence in the same quantization noise
environment, thereby eliminating the anomalous errors in the
music reproduction. These are the intrinsic reasons that the
Sabre DAC reproduces music so accurately

As you can see he directly references the dynamics within the complex music signal & that dealing with low level signal between two transients is one of the issue which is audible as "anomalous errors in the music reproduction"

Perhaps you should talk some more to your friend Mallinson about these issues?

 

[mmerrill99]

Yes, that rings a bell

So just to put some meat on the bones & why KSTR considers Diffmaker a toy - here's what he says is the correct way to do differential input to output measurement using music John Curl's Blowtorch preamplifier part III

And I would ask if Scott or BV are doing this or anything close to it?

The time-domain averaging only works when you have sample-sync'd recording while playback (DA and AD in the same device or even the same chip, running on one single clock), so analog sources and non-sync'ed sources (like two CD-players) are out.
I do it like this (basic procedure) :
- select a snippet of audio, about a second or so, for example a 100k sample block (select a block size and sample rate to get good reduction of mains hum+harmonics)
- loop this 1000++ times in a DAW, ad random dither to the whole length.
- record while playback through the trimmed analog diff and gain stage.
- throw away blocks at the beginning to avoid settling tails.
- condense the recording to one single block of samples, this is the averaging (wrote a piece of software to do that).
This gives a residual with at least 20dB less noise than a simple subtraction. The trimming is somewhat iterative once you get deep below the analog noise floor (no meaningful residual without heavy averaging).

As described, you can also record without the analog diff, sequentially, with and without DUT (or DUT A and DUT B) -- it is important to use the same DA and AD channel, hence not in real time using L/R channels.

It doesnt matter if you average first, then try to trim for lowest residual or vice versa, but averaging first is better because the improved SNR helps a lot to judge the residual.

Haven't played with diffmaker much but IIRC it does not do averaging (to do that it would need some input about the properties of the data it is processing, the block size). It offers resampling to take out some effect of non-synchrous recording by autocorrelation, so some jitter and/or clock drift is taken care of, but I could never get that to work satisfactory.
Diffmaker's trimming does work best with good preprocessing, that is pre-averaged, sample-sync'd and aligned input data.

 

[Cal Weldon]

The Moderation team is noting a good deal of repitition and run-on. We are closing the thread while we discuss its future.
Thank you for your understanding.