Diffuse Field Export

[Raccoon98172]

First of all, thanks for ASH-Toolset, it's a fantastic tool.

I've been working on using my personal HRTF measurements and achieving accurate headphone equalization and noticed that ASH-Toolset normalizes HRTF to the Diffuse Field. This makes it difficult to apply correct headphone equalization later, as the tonal balance changes.

I implemented a small workaround that exports the extracted Diffuse Field to a .txt file (REW compatible) during the processing stage. This allows users to restore Diffuse Field using an external equalizer (like Equalizer APO).

I've published the code and a guide on how to measure HRTF in this repository: https://github.com/Raccoon98172/generate_hrtf/wiki

Specifically, I added a save_diffuse_field_txt function and called it within hrir_processing.py.

It would be great if you could consider adding a native option to export this Diffuse Field data.

[ShanonPearce]

Hi,

if I understand the scenario correctly you would like to be able to skip the diffuse field calibration step in the HRIR processing so that the diffuse field (direction-independent) information is retained?

If so, would a simple toggle option for diffuse field compensation suffice? That way you can toggle off that step and keep the HRTFs as per original dataset.

If you would also like an option to export that curve as a filter I can integrate that as well but it may not be required if I include the above option.

[ShanonPearce]

I note that there are 2 stages of diffuse field calibration, 1 in the HRIR processing and 1 in the BRIR processing. I can implement an option to retain the diffuse field (direction-independent) information from the HRTF dataset all the way through to the binaural dataset output.

[Raccoon98172]

I'd like to clarify what I'm trying to achieve.

HRTFs are measured with microphones placed at the entrance of the ear canal. Therefore, when using headphones, we need to reproduce the same frequency response curve that was present during the HRTF recording.
To accomplish this, the headphones need to be equalized to a flat response at the ear canal entrance.
In other words: I measure the frequency response of the headphones using the same microphones (at the ear canal entrance) and invert it, creating the flat response I mentioned above.
Once the headphones are equalized flat, we can simply reproduce the same HRTF curve that was originally measured. In this scenario, the headphones no longer color the result because they now have a flat frequency response.

I don't fully understand the purpose of the second stage in the BRIR processing. I'm wondering whether preserving the diffuse field (direction-independent) information from the HRTF dataset could disrupt any processes within ASH-Toolset (Does it compensate for tonal changes introduced by reverb/sub integration?) and negatively affect the sound quality?

From my understanding, adding a simple toggle option to enable/disable diffuse field compensation should be sufficient.

[ShanonPearce]

Thanks for the context, this specific scenario is clear to me. Its slightly misaligned with the current processing that occurs in ASH-Toolset with regard to the diffuse-field calibration stages as it currently operates under the assumption that the headphones have also been diffuse-field calibrated but not equalised to a flat response at ear canal.

From my understanding, adding a simple toggle option to enable/disable diffuse field compensation should be sufficient.

An option to effectively nullify the HRTF related diffuse field calibration should make the result compatible with your headphone EQ method. For complete compatibility I will make it extract the diffuse-field curve at the initial HRIR processing stage and re-apply it at the end of the BRIR processing (if toggled), like with your current workaround. Ill include this option in the next release.

I don't fully understand the purpose of the second stage in the BRIR processing. I'm wondering whether preserving the diffuse field (direction-independent) information from the HRTF dataset could disrupt any processes within ASH-Toolset (Does it compensate for tonal changes introduced by reverb/sub integration?) and negatively affect the sound quality?

The second stage is simply there to remove any other unwanted direction-independent information from the reverberant response such as room related colouration, sub integration, measurement artefacts, etc. before applying any additional curves such as room target or headphone-ear compensation.

[Raccoon98172]

Thank you! Everything works great.

I'd like to ask one more question. Now theoretically I can make .sofa files with only 2 directions (30 and 330 degree azimuths) for my friends and use them in the program, so I don't have to record the full HRTF sphere? (Simple stereo for listening to music and movies)

There doesn't seem to be any audible difference.

[Raccoon98172]

It's probably acceptable though, as it's very hard to hear the difference.

[ShanonPearce]

I wouldn't expect there to be an audible difference between the 2 scenarios as the common transfer function (diffuse field) of the HRTF is reintegrated into the final response so in theory the result should be the same.

The differences might come down to how the direct sound (HRTF) integrates with the reverberation. A full sphere is reconstructed as an interim dataset for further processing which means even if you supply a SOFA dataset with only 2 directions it will try to fill the interim dataset with nearest available directions. That reconstructed sphere is used in the 2nd stage of the diffuse field correction which calibrates the full reverberant response before applying any final adjustments like room targets or HRTF diffuse field reintegration. That stage could result in slightly different equalisation if a different sphere of HRTFs is generated.

Either scenario should work fine but if you would like me to investigate further feel free to share both of your sofa files and I can have a look.

[Raccoon98172]

For the experiment, I used "VP_N6-ECEbl" from sofacoustics.org and a second file where I removed all points except 30 and 330 (I've attached them to the message).

This is a very interesting use case for ASH-Toolset because it significantly simplifies HRTF measurements.

VP_N6-ECEbl.zip

[Raccoon98172]

Hi! Happy New Year!

Any updates on comparing those two SOFA files (full vs 2-direction)?

[ShanonPearce]

Hi again, that scenario is a bit tricky but I did a deeper dive to see what can be done.

To recap: the inherent problem is the HRTF dataset needs to be diffuse field calibrated to ensure tonal compatibility with the reverberation response which is also diffuse field calibrated. Then the HRTF's common transfer function (CTF) can be optionally reintroduced at the end. That's fairly straight forward when a full sphere of directions is supplied but It becomes difficult when the HRTF dataset has few directions as the common transfer function is not really a diffuse field measurement in this case.

The solution I came up with was to dynamically generate an adjustment filter that is derived from an averaged listener. It calculates the difference between the expected DF curve (dense grid) and the actual DF curve for the directions supplied, using that listener as a reference. Once the CTF is adjusted the resulting HRTF more closely resembles the expected response. Refer to below plots for example.

Blue is the expected CTF curve (all directions), red is the CTF from 2 directions and green is the CTF from 2 directions but adjusted as per above. Its closer to the expected curve but still some deviation at the higher frequencies.

When looking at the calibrated HRTF results (DF compensation enabled) we can see an improvement in accuracy.

Then if we look at the final integrated response (with reverberation) we can also see an improvement. Bottom curves are DF calibrated whereas the top has retain DF activated. Blue is with the 2 direction dataset, yellow is with the adjustment added, and green is the reference curve with all directions.

[ShanonPearce]

I did another comparison between the integrated responses with different scenarios as the above had an error. When it comes to the retain diffuse-field case the results are fairly similar even with the adjustment. The biggest improvement is seen in the diffuse field calibrated case, the adjustment results are closer to the reference curve.

Bottom curves are DF calibrated whereas the top has retain DF activated. Blue is with the 2 direction dataset, yellow is with the adjustment added, and green is the reference curve with all directions.

[Raccoon98172]

Thank you for such a detailed explanation.
If we look at the comparison graph that I sent initially (with the red and blue curves), we can see that there are very few differences in the case of the ear that is on the side of the virtual speaker (ipsilateral), but there are differences in the case of the ear that is in the shadow (contralateral).
In that graph, the top two curves are on the side of the virtual source and they hardly differ at all, the curves that are 10-15 dB lower are the opposite side (in shadow) and you can see a small difference in some places.
Accordingly, the improvements should be in the case of the contralateral ear.

Based on the results, even without adjustment added, both should sound very, very close. This is good news for using a .sofa file with only 2 directions.