Generated acoustic spaces sound different in v4.0.0 than before

[ItsPasi]

I was using an acoustic space that a friend made that sounded much cleaner vs the provided spaces. He simply made a clean impulse response file so most room reverberation was cut out but still provided surround realism.
I will have the acoustic space attached.

Loading this acoustic space into the new v4.0.0 update changed the sound in a significant way and in my opinion it doesn't sound as clean and realistic anymore, and the acoustic spaces made in v3.7.0 won't load anymore.

I tried both FABIAN and KU-100 listeners in the space generation (seems to be the main new option in the menu) but both sound less precise than the old acoustic space. The only value I normally change when generating these spaces is the Rise Time which I set to 1.0.

Flat Room Response.zip
I put the folder inside the zip in the acoustic space input folder and generated the space afterwards.

(also not sure if only this space is affected or if every space sounds different now, but I definitely hear a significant sound change that feels less precise)

[ShanonPearce]

Hi,

I am planning to upload a new release (hopefully within the next couple of weeks) which will have an updated acoustic space import tool.

• Fixed a small bug impacting time domain alignment accuracy that was in the latest release.
• Some more parameters that were previously hidden will be exposed. This will allow finer control without me having to tweak the code in future.
• Default parameters will be set to align with the defaults used in v3.7.0 so the quality should closely resemble that version

Preview:

In the meantime if you would like to use your previously imported acoustic space, you can generate a new one and then copy across the "dataset_-NAME-.npy" file from the previous version and replace the one that it generated.

[ItsPasi]

Okay I tried out the new version now and I think I realize a few things being different still. First of, (and this was happening in the old version too) generated rooms will sound significantly different after every new creation, even when the settings stay identical.

Then I noticed when I replace the file as you recommended, I still hear a difference in the new version vs when I play it in 3.7.0. I am not sure what is different, It sounds less focused but it might be that it simply has a slightly different EQ. Was there something that changed in that regard?

Also, newly generated spaces in 4.1 still sound different than before, probably outside the general creation to creation difference.

I have provided a sound demo that demonstrates these differences. This should show the difference in sound quiet clearly.
If I had to describe it, the acoustic space, when I play it in 3.7.0 sounds the most precise and focused. When I play the same space in 4.1.0 (the way you explained it) it sounds slightly less focused. The generated spaces in 4.1.0 also have a weird double/echo sound in the bass that you can maybe hear in the video.

sound.difference.mp4

[ShanonPearce]

1. The generated rooms will sound slightly different after each creation due to the binaural conversion process using a random distribution to simulate directions on the spherical array. If consistency is required I can include a seed parameter so that the same distribution can be used for each creation. The results will then be identical.
2. Differences in EQ between the 2 versions using the same acoustic space likely comes down to a) the 'Headphone Compensation' parameter which changed slightly but will only impact mid/high frequencies, b) calibration octave smoothing parameter in advanced settings which can be tweaked. If you can provide a copy of the WAV outputs from both versions or the acoustic space output that was generated in 3.7.0 I can probably identify where the difference lies.
3. The bass issue might be caused by the 'Decay Curve correction' parameter - try toggling that off as it was disabled in 3.7.0. Also try increasing the spherical grid density as higher values usually result in smoother responses.

[ShanonPearce]

Regarding point 1, I will include a distribution seed option in the next release to allow for consistent acoustic space creation.
For point 2, I will include the filters from v3.7.0 in the 'headphone compensation' list so that the same EQ can be achieved.

[ItsPasi]

1. I would greatly appreciate a seed option in the creation.

2. I will send the WAV files I use (30°) here:
   BRIR_3.7.0_A30.wav
   BRIR_3.7.0_A-30.wav
   BRIR_4.1.0_A30.wav
   BRIR_4.1.0_A-30.wav
   Also, if it truly is just a difference in EQ, then there is no need to provide the old settings. I am just unsure if it truly is, but EQ is something that can be personally tweaked, so no need to add unnecessary options for that.

3. I tried turning off the Decay Curve Correction, but it still feels like there is some timing problem going on, but I can't say for sure, it just doesn't sound nearly as focused vs to old creations. I also tried tweaking the settings in the creation a bit and couldn't really find a single one that improved the output, so not sure what is going on there exactly.

[ShanonPearce]

Thanks for providing the sample outputs, I can see some subtle differences. To help investigate this a bit further would you also be able to provide:
a) the input IRs used to create the acoustic space (ash-toolset\data\user\IRs)
b) the resulting imported acoustic space files from both 3.7.0 and 4.1.0 (ash-toolset\data\interim\reverberation\user)
c) the exact SOFA file you used for listener selection

It likely comes down to a parameter change but Id like to verify by replicating
a) the acoustic space result from 3.7.0 (or close approximation) using the same input IRs
b) the BRIR outputs from 3.7.0 using the same acoustic space you previously generated and same listener

[ItsPasi]

Audio Debug Files.zip

I hope the file naming structure won't be too confusing. I exported a bunch of stuff from scratch and generated all of this new. I used 2 different acoustic spaces, but the 32_bit folder is what I used before to generate new spaces.

I included both spaces generated in 3.7.0, 4.1.0 and ones I converted from 3.7.0 to 4.1.0 (all have specific naming)
Spaces generated in 4.1.0 sound significantly worse (with he weird timing issue). I also only included the best sounding outputs, I generated many more and just kept the cleanest sounding ones.

[ShanonPearce]

Thanks for the data. I reviewed the acoustic space functions and parameters alongside your outputs for mp_clean_impulse_response and compared the results. I managed to resolve the timing issue (caused by excessive delay) and improved the overall quality (via slight filtering tweaks) however I was unable to replicate the exact result from v3.7.0 for mp_clean_impulse_response.

Ill continue to dig a bit further as the 3.7.0 result has some additional delays in the high frequencies resulting in a more externalised sound but for now Ill do another minor release soon to address the main issues.

• New acoustic space parameters relating to time domain alignment are included to resolve timing issues
• New seed parameter allowing for consistent results. Will default to the same seed used in 3.7.0.
• The default acoustic space parameters will result in the same timing/delay as in 3.7.0.
• Some new HRTF filtering and EQ options are also added to allow for binaural sim wav outputs to more closely align with 3.7.0.

[ShanonPearce]

If you come across any timing issues in the most recent release (v4.2.0), try reducing 'Search Offset (samples)' to 0. This will minimise delays, causing reflections to arrive earlier.

[ItsPasi]

Okay, I tried the new version now and honestly at least I notice no real improvement in any of the points I brought up. Additionally now the 32_bit folder I used to generate acoustic spaces isn't working as an input anymore.

1. For me when I load one of the old generated spaces they still sound less clear than in 3.7.0, even when I use the 3.7.0 compensation filters. I don't even think the compensation filters (same type, different version) make any noticeable difference in the sound. And I also honestly don't think they should exist as this just adds too many options to choose from.

2. The newly generated spaces seem to have the same timing problem, I tried reducing the value as you suggested but that made it worse. Also if I didn't mention it, the timing problem seems to be caused by my left and right channel to be delayed from each other. It's messing with the center balance and is the clearest with bass hits that sound doubled.

3. While I know this was probably not planned to work anyway, the 32_bit folder was an output space but could also be used as a new input space before, which I used and I preferred the sound somewhat. But as long as I can just use the old generated space in the new versions this doesn't need to be included yet again either (low priority in my eyes).

[ShanonPearce]

1. I was able to replicate the same output for the mp_clean_impulse_response result from 3.7.0.

• In advanced settings, untick HRTF->Zero-Phase Suppression Filters. Then under User Data, clear the cache by clicking Maintenance->Clear Local Data Cache (this forces the app to prepare new HRTF data).
• I also unticked LF Extension->Zero-Phase extension filters as that was disabled in your example (previously called Forward-Backward Filtering).
• I set the other parameters as per the name and the resulting WAV output mostly aligned with the provided WAV: 3.7.0BRIR_E0_A-30_EarlyReflDel0.8_MirrorRightSide_DirectSoundGain0_FlatRoom_Flat_3.7.0_mp_clean_impulse_response.wav.

2. I'll compare the new acoustic space outputs vs the ones you provided again to see if I can find the exact issue you are describing. For now some other parameters worth tweaking are min. shift limit and max. shift limit. Previously in v3.7.0 they were set to -4.5 and +1.1.

3. The latest version will auto detect if an acoustic space output is placed in an input folder and will copy it across to the output folder instead of processing it. To get around this you can delete the metadata csv file from the input folder which the app looks for to detect npy files that are already processed. Once the csv is deleted the folder will be treated like a regular npy dataset input.

[ItsPasi]

Okay, I got it to sound identical now too. I forgot to apply the 3.7.0 loudness compensation (I guess this was what made the biggest difference in the end). But since the differences here are actually just EQ then I assume, there really is no need to keep this many options in the program, especially because any future change would make it just more complicated in my opinion.

Third point is solved then as well.

I guess only the space creation has any issues still, not sure why.

[ShanonPearce]

The newly generated spaces seem to have the same timing problem

Regarding this point, I was able to generate an output very close to the one you provided from 3.7 (Flat_3.7.0_mp_clean_impulse_response). I had to reintroduce the previous spatial expansion method I used in v3.7.0 which used a different filter creation algorithm. Ill upload a new incremental release that includes this option in the acoustic space tool and will provide details on which parameters to select.

But since the differences here are actually just EQ then I assume, there really is no need to keep this many options in the program

I will keep the recently added legacy EQ options hidden by default but can be made visible if desired.

[ItsPasi]

Thank you for trying so hard to solve the issue. I am also willing to try pre-release versions to check before having to publish a new release.
Also, honestly the goal wouldn't even be to match the old sound exactly, the main issue that needs to be solved is the timing offset issues. It leads to my outputs sounding like bass hits are doubled and sometimes even breaks the imaging, like the frontal projection that normally happens (with a fitting HRTF for my ears) breaks.

If the old algorithm fixes that then that would be awesome.

[ShanonPearce]

In the newest version (4.2.1), if you set alignment method to 'standard' and spatial expansion method to 'Pitch shift - simple', the result should be similar to what was rendered in 3.7.0. I also noticed there was a peak around 20khz which can be removed by increasing the calibration octave smoothing value and calibration high cutoff freq in advanced settings.
That should allow for results similar to earlier versions.

As for the general timing offset problem, I will have to investigate it a bit further if it can be tightly controlled.
Some things worth trying:
Acoustic space import -> alignment method set to custom and shift limits set to -6 and 0 to minimise delays. Search offset around 200. If offset is reduced to 0 it will minimise offset but could distort the bass.
Advanced settings -> clear local data cache, then increase suppression cutoff frequency to around 200hz. This will reduce HRTF energy in the bass as an experiment to analyse the impact. The double bass hits could be from the direct sound (HRTF) immediately followed by a strong early reflection introduced by the acoustic space.