davircarvalho
commented
1 year ago Hello @ABoredBunny, thanks for your interest in this repository and your feedback! I'll try to answer your questions and clarify some points as I can.
This solution only outputs one field it doesn't produce fields that can be used for near field synthesis.
And it only outputs a set field with specific density.I will assume what you are referring by "field" concerns the measurement positions or the sampling grid (if it's not it, please forgive me) -- There's a very specific and boring reason why this sampling grid was chosen, and it's related to the datasets used to train this model (there's a whole paper about this under peer-review, so I can't share it with you right now). But what really matters is that this repository already contains an interpolation function you can use to generate HRTFs for whatever direction you want. This function has at least 4 interpolation methods to explore, go nuts. But keep in mind the output grid from the app will not work well with spherical harmonics interpolation or extrapolation, and bilinear interpolation will not work for positions beyond the the elevation limits of this grid, other than that, this function will let you output the relevant grid for most use-cases.
About near-field extrapolation, I've invested a lot of time doing it for the auralization app . But honestly, I have no motivation to bring that to the individualization app. Although if you have HRTFs that cover the full sphere (not the ones from the app) you could simply use the other function I mentioned above, for interpolation, set the flag for spherical harmonics and provide the azimuths, elevations and distances you desire (just check if it's using the ITA api, it's the only one that includes distance with spherical harmonics). AFAIK that should still be state-of-the-art. If you need help generating a grid, a good starting point is this function.
but you should probably mention that these HRTF files are meant to be used together with Hesuvi HP filters. Because they sound wrong with other harman target equalization because they EQ after 10KHz.Sorry, I honestly don't know what you mean by that, HRTF spectra above 10kHz contain the monoaural cues for localization related to the pinna shape. Hesuvi HP filters, I assume are headphone correction filters, they are supposed to reduce the influence of the headphone itself in the auralization process, that's a totally independent procedure, that has a separate physical meaning from HRTFs. Harman target equalization is the only EQ per se, as it equalizes the content to have a balance that "most people like". So ... I'm not sure what's your point :)
ABoredBunny
I first of all want to thank you for making this, this is the most convenient and user friendly way of making a personalized HRTF yet. Going from an anechoic chamber with speakers. To needing a dense 3d model of ones head and a decent computer. To only needing a caliper and about a minute of compute time, has been quite the trip.
But this solution still has problems: This solution only outputs one field it doesn't produce fields that can be used for near field synthesis. And it only outputs a set field with specific density.
For the first one, I suggest this paper here's someones implementation in matlab. This approximation creates pretty convincing results, it is implented in Binaurilizer_nf. Something like a field where you can type in distance and generate a field, or specify multiple fields during generation.
The second one is up to how you may implement it, what I would want is the ability to interpolate the points to an arbritrary grid. Possibly using a config file which specifies the number of rows and there spacing in degrees or similar. Because some things don't have there own interpolation, Or expect certain grid types namely OpenAL Soft.
Also, somewhat unrelated, but you should probably mention that these HRTF files are meant to be used together with Hesuvi HP filters. Because they sound wrong with other harman target equalization because they EQ after 10KHz.