Question about the Task 1 Stage 1

[Yutong Wen]

Hi SONICOM LAP Challenge organizers!

I'm participating the task 1 of this challenge. Thanks for organizing! Our group currently has troubles passing the stage 1 evaluation. The rmsL, rmsP, and querr thresholds are not passed. So we did some experiments with the ground truth hrtfs. We found that with a simple scaling of the real hrtf would fail the stage 1 evaluation. Specifically, we processed a person's hrtfs just by dividing the max value. And the thresholds that were not passed are also rmsL, rmsP, and querr. In order to perform well in the stage 2, we think it's inevitable to incorporate some kinds of scaling to the original hrtfs, so we wonder how to deal with this situation. And we also wonder how the baseline method deals with this. Thanks!

Best,

Yutong

[Aidan Hogg]

Hi Yutong,

Sorry for the delay.

We assume you are referring to “overall” amplitude scaling here?

In the threshold computation, each sofa file is normalised (i.e. the amplitude of the HRIR is divided using the maximum absolute amplitude found in the frontal position - barumerli2023_featureextraction line 172, 174) to get values between [-1, 1]. This is done separately for the original and harmonised SOFA files. Therefore, an overall amplitude increase (or decrease) should not be a problem.

However, this would be a different story if the amplitude ratios between the HRIRs were changed in the harmonised SOFA file. The model would detect this difference with the template comparison procedure.

We are not entirely sure what the problem you are having is. Would you be willing to share your code (privately, if you would prefer) so that we can reproduce it?

Cheers,
Aidan, along with the rest of the LAP team

---------------------------------------------------------

Dr Aidan Hogg

Lecturer at Queen Mary University of London

Honorary Research Associate at Imperial College London

Centre for Digital Music 

Electronic Engineering and Computer Science

Queen Mary University of London

327 Mile End Road, London

E1 4NS, U.K

Email: a.h...@qmul.ac.uk

Personal Website: aidanhogg.uk

QMUL Group Website: c4dm.eecs.qmul.ac.uk

QMUL Website: qmul.ac.uk/eecs/people/profiles/aidanhogg

Imperial Group Website: axdesign.co.uk

Imperial Website: imperial.ac.uk/people/a.hogg 

[Yutong Wen]

Hi Aidan and the LAP team,

Thanks for your reply! Yes, we're willing to share our sanity check code! The code includes the sofa read and write functions and a simple processing code for normalization in the either time or frequency domain (yes, amplitude scaling). We're not sure the sofa file writing process is entirely correct, but we at least make sure that the written normalized hrtfs have the right normalized amplitudes and correct corresponding source locations. I emailed to you the code and the sample written normalized sofa file and its original sofa file, as I wasn't able to post them here due to the system kept saying my post is was long. We really appreciate your help!

Best,

Yutong

[Aidan Hogg]

Hi Yutong,

Thanks again for sending your sofa files. We have double-checked, and we are now pretty sure that you are storing the sofa files incorrectly; I can get my code working with the same operation.

Could you double-check your code or perhaps try using a different sofa file package? There are quite a few of them. I know ‘pysofaconventions’ is quite popular, and some prefer just to use
‘netCDF4’ directly.

I actually currently use sofar in a similar way to you (see below)… Maybe you can also check the coordinate system you are using...

import sofar as sf

sofa = sf.Sofa("SimpleFreeFieldHRIR")
sofa.Data_IR = full_hrirs
sofa.Data_SamplingRate = hrir_samplerate
sofa.Data_Delay = delays
sofa.SourcePosition = source_positions
sf.write_sofa(sofa_path_output, sofa)

I really hope you can find the issue and resolve it quickly.

[Yutong Wen]

Hi Aidan and the Lap team,

Thanks for your reply, and thank you for sharing the package with me. We will try to use another sofa package to write sofa files. We appreciate your help!

Best,

Yutong

--
The IEEE Signal Processing Society is sponsoring the Listener Acoustic Personalisation Challenge.
---
You received this message because you are subscribed to a topic in the Google Groups "SONICOM LAP Challenge" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/sonicom-lap-challenge/6MmsmqAymQg/unsubscribe.
To unsubscribe from this group and all its topics, send an email to sonicom-lap-chal...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/sonicom-lap-challenge/02c93831-e178-4847-a7bd-90841c8cf6e1n%40googlegroups.com.