Neutron EQ Accuracy

[MotleyG]

Okay, so they played a sweep ( constantly rising tone ) on a smartphone, with several equalizers set to a common preset, fed the smartphone's analog output ( or the digital one ? ) into a "Creative GC7" ( an external DAC ? ) and recorded it there, then imported the data into "REW / Room EQ Wizard" ( audio software ). As reference the REW calculated the output from a flat input and the eq preset. Did I get that right ?

Unless I missed it, there was no device make/model listed. There is mention of a Samsung USB-C "dongle" and that it is 1.0Vrms capable but limited to 0.2Vrms typically. Seems this could have limited audio quality and not fit into the true "hi-res" category. I'm doubting the 192kHz capability claim as well for the same reason. But obviously this means the phone DAC's are bypassed, and the signal is being converted to analog on the way to the outboard device measuring the response.

The test signal was generated from the test software, 10Hz-22kHz at 48kHz (16 or 24 bit?). So this limits the frequency response testing already.

The tests for both Wavelet and Poweramp were run at 44.1kHz and 192kHz. Likely this was influenced by the 48kHz sample rate test tone that was generated and possibly had some effect on the stepping results in those measured curves. The additional results measured afterwards with Neutron may have been coloured by measurements run at different times. And certainly if 64-bit processing was enabled, that too could account for the extended high frequency improvement I see runnig out to 20kHz and beyond.

The curves themselves are all very close across most of the frequency range. Seems there is a small level difference for each, that could be simply caused by unmatched preamp levels, Replaygain, etc. in the software of each app. If these were dynamically matched (adjusted to overlay each other without the volume offsets), it would be more clear to see where the true response differences stand out.

This was not the most scientific or controlled test scenario. Certainly the results and the interpretations were subjective. Just as much as my personal comments are here. Without having all of the other subsets of data and the options that were enabled or disabled in each of the tested apps, there is no way to fairly assess where these identified gaps have come from.

[dmitrykos]

Thank you very much for pointing to this issue!

After careful check of the implementation I can confirm that there was imprecise conversion of Q for Shelf filters that resulted in less steep slope which was correctly noticed on Reddit. The correct steep slope can be achieved with S parameter set to 1. So if you are using S parameter for Shelf filters then you are not affected, issue happens only if Q is used as parameter. FRC is affected too because AutoEQ presets are using Q for Shelf filters. The problem concerns only Shelf filters, all others are ok.

In general this problem is hardly noticeable by ear but of course instrumentally it can be evaluated that was demonstrated in Reddit's post.

I noticed mentioning of 64-bit processing by MotleyG and can confirm that all filter-related arithmetic is done in true 64-bit fp resolution - coefficients calculation, as well as processing of the sound data by the filter.

Next public update will contain this fix.

[blaubär]

2.22.3 (08.2023):
[...]
! Fixed:
- Filter DSP: Q giving insufficiently steep slope for Shelf filters (HS, LS)
[...]

[fankyy]

In order that EQ settings are applicable cross-application and cross-platform wise, it's good this was fixed of course, but: If their measurements are accurate, the corrected implementation might lead to certain frequencies exceeding 1dB deviation. For some headphone models I use a low shelf boost up to 15db, there it might deviate even more and lead to upper-mid bass sounding harsh (or in the opposite slightly too thin) also for some high shelf corrections.

I've a copious amount of presets for a variety of headphones and speakers and adapting each preset to the new implementation by ear would become quite an overwhelming task. So what does it mean if I update to 2.22.3:

- will the existing parameters of every preset get adjusted automatically so they'll sound the same or
- will there be an option to switch between the old and new implementation? or
- at least such a feature planned for a future release?

[blaubär]

So what does it mean if I update to 2.22.3:

- will the existing parameters of every preset get adjusted automatically so they'll sound the same or
- will there be an option to switch between the old and new implementation? or
- at least such a feature planned for a future release?

- No, the presets will stay as they are, and the new version will correctly implement them
- Therefore the sound will slightly change for shelf bands of typ Q
- There is no switch present or planned
- The change will be predictable

I suggest you ask neutronmp@gmail.com for a formula to calculate the new preset value from the old one to keep the sound as it is, so that you can change your presets easily.

[blaubär]

I just installed the new version on my pc, where I had some shelf eq bands which I wanted to look like S=1, but which were of type Q. After the update the curve had changed, I changed type to S and set S=1. The curve then looked as it should and had done before. I then changed back to Q, which then had a value Q=0.707 .

So IF this is linear THEN you just have to multiply the value of Q by 0.707 . Mind, big IF .

[whatsisname]

hi just transferring my neutron setup from my old phone to my new one. old phone has the non generic install version so I can't just copy the folder over, got to recreate eq presets again manually. new install is the generic install version.

on the low and high shelf eq settings, when I match the Q value numerically, it looks very different on the graph. old setup is version 2.19, new one is version 2.22.

2 questions come to mind.

- why is this?

- which is correct, the numerical figure or the graph. which should I copy to correctly match my old setup?

the peak bands don't have this problem. graph and numerical figure match old device.

any clues? perplexed!

[WitzyZed]

Build 2.22.3 (current build is 2.22.5)

Code: Select all

! Fixed:
 - Filter DSP: Q giving insufficiently steep slope for Shelf filters (HS, LS)

[whatsisname]

I have 2.22.5. so it's build 2.22.3 that has the fix but it's broken again in 2.22.5?

[WitzyZed]

As to why there was behavior differences between 2.19.X and current.