Another Spot The Difference Thread

[charleski]

Yes, it's yet another of those 'can you hear a difference' threads! I know you all love them!
This one even has a special Dr Evil twist to it!

The music is a segment taken from Frode Fjellheim: Kyrie by Cantus & Frode Fjellheim published by 2L and made available on their test bench site. Since it's only a 1 minute segment I think it should be OK to post here.

Try to decide if you can hear a difference first before poking around with analysis tools, but I encourage taking a look after you've made your choice. The files are precisely matched for loudness to 0.01LUFS.

The files can be downloaded at
https://drive.google.com/file/d/1M07bc5Ubv4D8PyXC7SDlBorKiOt2ysG8/view?usp=sharing

 

[Helicopter]

I am not sure, but I at least convinced myself b might have more and better sub-bass. The cymbals sound compressed and klinky on both recordings.

 

[dasdoing]

I vote "no"...with the desclaimer that I am against "difference hunting". If there is no obvious diference, I wont listen a 100 times to desperatly find one

 

[JeffS7444]

I was in the mood for this sort of thing, so I found myself an ABX comparator (Lacinato) and without looking at file sizes performing spectral analysis etc, loaded them up and did an ABX.

Spoiler: What I (thought I) heard

Assuming that I used the software correctly (it shuffles position of the files) I felt that "Kyrie A.flac" had higher-fidelity sonics, both deeper bass and much better sense of "air". I wondered if "B" had been bandwidth-limited, with tambourine sounding less distinct.

Equipment used: Lenovo Thinkcentre 91Z, Objective DAC (original), Topping VX3 power amplifier, Overnight Sensations speakers. Setup had been previously eq'd with the aid of REW.

 

[Maxicut]

I couldn't be bothered trying to pick the difference by listening, so I loaded both files into Pro-Tools & they didn't even come close to nulling.

 

[Helicopter]

I couldn't be bothered trying to pick the difference by listening, so I loaded both files into Pro-Tools & they didn't even come close to nulling.

 

[dasdoing]

OK, yesterday I listened on my TV speakers which go down to 80Hz or so.
now listening on my monitors it seams to me as if version b is high passed

 

[stalepie2]

Interesting music... don't think I can hear a difference, though.

 

[charleski]

Hardly anyone tried this, pity.

Kyrie B is the result when Kyrie A is phase rotated +90 degrees. There is a 1dB difference at 20Hz, which fades to 0 from 35Hz up. This is a side-effect of the processing on this particular segment of music and frankly I doubt that is enough to be discernible unless you have a very good bass setup, but maybe I should have tweaked the segment chosen so their spectra matched more closely.

Phase rotation is completely different to phase shift. There's an introduction to I/Q representation (which is ultimately just about how a circle turns into a sine wave) here and an introductory discussion of phase rotation here. While phase modulation and its digital equivalent, Phase-shift keying, is heavily used in radio communications, you don't come across it much in audio. But it can be very useful for correcting the asymmetrical waveforms often found in recording vocals (particularly plosives) and percussion - in such circumstances it's common to find that one side of the waveform has much greater excursion than the other, even though there's no overall DC shift. One solution is simply to compress it, and since that requires almost no effort, it's commonly used. But a more sophisticated answer is to record with high dynamic range and keep the dynamics intact but then rotate the phase to equalise the positive and negative halves of the waveform before dithering the mix down to 16bits for delivery. Most engineers regard this as having no audible effect. The phase rotation was performed in Izotope RX8.

Even though they sound the same, attempts to null the waveforms will fail if you are only aligning the real parts, since the quadrature components need to be aligned as well. So the PKMetric produced by DeltaWave fails and produces an RMS value of -23.7dBFS. I tried a range of different rotations and noticed that while the DFMetric median from DeltaWave increased monotonically as the phase of the compared sample was rotated from 0 to 90 degrees, the PKMetric results were all over the place and showed no correlation with degree of phase shift. Since phase modulation is so rare in the audio realm I don't know if this is of any significance though.

 

[dasdoing]

There is a 1dB difference at 20Hz, which fades to 0 from 35Hz up.

so I was right?

I heard it with the strong bass note around 39 seconds. it begins earlier, but around 39 seconds it is strongest

 

[charleski]

so I was right?

I heard it with the strong bass note around 39 seconds. it begins earlier, but around 39 seconds it is strongest

There's a strong bass note at 39 seconds, but it's 73Hz.
Here's a FFT of the two seconds from 38.0 to 40.0 with Kyrie A in red and Kyrie B (rotated) in green. At 73Hz they have identical intensities, but as you can see there's a tiny difference at 20Hz.

 

[dasdoing]

sorry for late reply,
I have no clue if I could ABX this or if I just was lucky - you graph sugest I shouldn't
BUT:
phase rotation of +90 degrees creates diferent delays at diferent frequencies.
the main diference will be therefore in the group delay.
at 22Hz (which seams to have some energy in your graph) we have 11.363636363636363ms of delay f.e.
while at 2200HZ we have 0.11363636363636365ms (basicly zero).

 

[dasdoing]

I think, you basicly aplied an all pass filter

 

[pkane]

Even though they sound the same, attempts to null the waveforms will fail if you are only aligning the real parts, since the quadrature components need to be aligned as well. So the PKMetric produced by DeltaWave fails and produces an RMS value of -23.7dBFS. I tried a range of different rotations and noticed that while the DFMetric median from DeltaWave increased monotonically as the phase of the compared sample was rotated from 0 to 90 degrees, the PKMetric results were all over the place and showed no correlation with degree of phase shift. Since phase modulation is so rare in the audio realm I don't know if this is of any significance though.

You can't use the simple waveform alignment routine in DeltaWave with the phase being so corrupted. The default settings in DW will deal with constant group delay and linear level differences. Anything else will certainly produce a larger error result. To correct for such complex phase differences, you should engage the non-linear EQ in DW. Here's the result with these two files:

Settings to reproduce this:

 

[charleski]

You can't use the simple waveform alignment routine in DeltaWave with the phase being so corrupted. The default settings in DW will deal with constant group delay and linear level differences. Anything else will certainly produce a larger error result. To correct for such complex phase differences, you should engage the non-linear EQ in DW. Here's the result with these two files:

The phase isn't corrupted, and the difference is only 'complex' in the sense that it's a manipulation of the complex expression of a sine wave. It's a very simple constant 90 degree phase rotation.

It looks like the non-linear EQ is trying to approximate the phase rotation by applying differential group delays to different frequency bands.

It's an interesting tactic, though it can run into problems. I mixed together a 200Hz sine at 0deg, a 1kHz sine at 45deg and a 15kHz sine at 90deg, then rotated the mix +90 degrees twice. Rotating them by 180 degrees in RX8 results in a simple polarity inversion, whereas doing 90 degrees twice introduces boundary effects at the start and end and adds in a microscopic amount of overflow error. If I mix this back with the original in Audition and trim off 80msec at the start and end to get rid of the boundary effects it nulls almost perfectly, with an average RMS of -138.9dB.

Analysing this in DeltaWave using the non-linear EQ setting given above produces this: (it asked to turn on simple waveform processing, so I let it, the results were almost identical without)

If I turned the non-linear EQ off, then DeltaWave got much better with a near-perfect match apart from the boundary effects:

So the non-linear EQ seems a bit hit-and-miss, which is not really surprising.

Interestingly, DeltaWave seems to work much better on the files I provided if you simply turn Correct Clock Drift and Subsample Offset off (I had both on before). (rest of the settings here are default as installed)

It seems that the attempts to correct the misaligned quadrature components by shifting the in-phase ones (which is doomed to failure) ended up making things worse. Of course, in practice you're going to keep drift correction turned on because it's almost certain that two signals from two different sources are going to have some sort of relative real drift.

But I think the most important consideration is that this is little more than a curio, as it's very unlikely that anyone will encounter phase-rotated audio signals in practice, unless there's something very wrong with the circuitry.