Skip to main content

Timeline for HRIR interpolation using VBAP

Current License: CC BY-SA 4.0

17 events
when toggle format what by license comment
Jan 19, 2021 at 12:53 vote accept Mattia Surricchio
Jan 19, 2021 at 12:53 comment added Mattia Surricchio Thank you, this makes everything clear. So basically the paper the states that it is enough to apply the VBAP directly to the original HRIR(ime.usp.br/~mqz/TwoApproachesForHRTFInterpolation.pdf) (which basically creates a weighted sum of the HRIR over a triangular area) is wrong? This would explain why the proposed method doesn't work and also why i'm getting the strange behaviours i mentioned in the original question
Jan 19, 2021 at 2:14 comment added ZR Han @MattiaSurricchio I've updated my answer about why should we perform the interpolation on the minimum-phase parts of HRIRs. Hope it's helpful.
Jan 19, 2021 at 2:12 history edited ZR Han CC BY-SA 4.0
deleted 58 characters in body
Jan 18, 2021 at 10:44 comment added Mattia Surricchio For example, in the above paper, they do not mention any conversion. Also in this paper: ime.usp.br/~mqz/TwoApproachesForHRTFInterpolation.pdf they seem to apply VBAP directly to the available HRIR. While in this paper: ieeexplore.ieee.org/document/7079779 they say "In order to avoid undesired destructive interference, the inter- polation procedure given by equation (1) should be performed on the minimum-phase versions of the measured HRIRs". I am quite confused on how to perform a decent interpolation of my HRIR
Jan 18, 2021 at 10:40 comment added Mattia Surricchio I changed my interpolation method to a simpler one (linear interpolation on the azimuth) as reported in umiacs.umd.edu/labs/cvl/pirl/vikas/projects/dynspatial.pdf . However I am getting audible artifacts such as a low pass filtering behaviour in the interpolated HRIR. Do you have any resources on how to perform HRIR interpolation? I checked many papers and I am quite confused by the large amount of available methods. I read about converting the HRIR to minimum phase ones in order to do interpolation, but i didn't understand if it's mandatory
Jan 18, 2021 at 1:29 comment added ZR Han @MattiaSurricchio Well, a big oops. Check your database and make sure that you have interpolated HRIRs in the right way. Generally it's not a hard thing and you have to keep careful.
Jan 15, 2021 at 14:48 comment added Mattia Surricchio Anyway I analyzed my code and data, and those spikes are the attack of kick and the snare in the song! I changed the input file with a piano song (no drums) and those spikes are missing. As you might notice they are also quite regular and i'm quite sure that the part where they're missing (in your plots), is the pre-chorus where the drums are missing
Jan 14, 2021 at 13:33 comment added Mattia Surricchio Thank you for your help! Would it be a problem if we continue the discussion via chat? I would like to know your opinion on my approach, since you are for sure more experienced than me on these topics
Jan 14, 2021 at 13:23 comment added ZR Han @MattiaSurricchio A decently working HRTF processing will not produce these peaks, unless your input signals contain these peaks periodically. Otherwise, there is something wrong with your framing operation. BTW, the database I use is Priceton's 3D3A database, which has 72 azimuths: [0°, 5°, 10°, …, 355°] and 9 elevations: [–57°, –30°, –15°, 0°, 15°, 30°, 45°, 60°, 75°] (72 × 9 = 648 positions in total).
Jan 14, 2021 at 13:11 comment added Mattia Surricchio If you are referring to those spikes, as you might notice they're almost periodical, they are probably the consequence of my real time processing. I'm basically taking n samples from real time audio and then i convolve them with the selected HRIR.
Jan 14, 2021 at 13:05 history edited ZR Han CC BY-SA 4.0
added 320 characters in body
Jan 14, 2021 at 12:48 comment added Mattia Surricchio Did you plot the waveform of the audio file? I actually didn't, but from the only listening perspective, it seemed that the volume of my audio file in the different spatial positions is more or less constant, there are no huge jumps in the perceived audio volume. Do you have any other database you could suggest me? The ones that i know (KEMAR, CIPIC ecc...) they still have the same problem i mentioned, you have fixed couples of (azimuth,elevation). Maybe i'm doing something wrong, but ordering azimuth and elevations indipendetly, could lead to the same problem i've already mentioned
Jan 14, 2021 at 12:25 comment added ZR Han @MattiaSurricchio Yes my approach to find triangles only works for mesh grid of azimuth-elevation pairs, so it doesn't work for your database. Maybe you can search for a mature VBAP code, or choose another database. I think most databases have uniformly distributed source positions, because it is easy to do the measurement. And the second point is saying that as seen from the waveform of your first audio example without interpolation, the jumping gains are also observed.
Jan 14, 2021 at 9:50 comment added Mattia Surricchio Are you assuming to order the azimuths and elevations indipendently? Because (maybe i'm wrong) if i do this way, i could end up with a couple of (azimuth,elevation) which doesnt exist in my dataset. I edited my question with an example of the available positions data. I am sorry but I didn't understand the second point, what do you mean by 'also found in the first case'?
Jan 14, 2021 at 2:49 history edited ZR Han CC BY-SA 4.0
edited body
Jan 14, 2021 at 2:24 history answered ZR Han CC BY-SA 4.0