An efficient binaural spatializer

The "binauralModeled-ircam" object is a pure-vanilla implementation of the binaural model created at IRCAM.. An example of its capabilities can be found at this page.

The object is found at my repository: https://github.com/andresbrocco/binauralModeled-pd

Main concept

Basically, it applies ITD and approximates the HRTF to a series of biquad filters, whose coefficients are avaliable here.

Sample Rate limitation: those coefficients work for audio at 44100Hz only!

Space interpolation

There is no interpolation in space (between datapoints): the chosen set of coefficients for the HRTF is the closest datapoint to the given azimuth and elevation (by euclidean distance).

Time interpolation

There is interpolation in time (so that a moving source sound smooth): two binauralModels run concurrently, and the transition is made by alternating which one to use (previous/current). That transition occurs in 20ms, whenever a new location is received.

Interface

You can control the Azimuth and Elevation through the interface, or pass them as argument to the first inlet.

Performace

Obs.: If the Azimuth and elevation does not match exactly the coordinates of a point in the dataset of HRTFs, the object will perform a search by distance, which is not optimal. Therefore, if this object is embedded in a higher level application and you are concerned about performance, you should implement a k-d tree search in order to find the exact datapoint before passing it to the "binauralModeled-ircam" object.

Ah, maybe this statement is obvious, but: it only works with headphones!

Thanks for sharing this! BTW, is the [!-~ 1] object inside the [pd crossfader] subpatch a typo? Should this be [!=~ 1] instead?

Hey @alexandros. that's not a typo! That's short for [expr(1-$f1)]: the exclamation point reverses the order of the inlets when preceding basic math operators + - * /

@andresbrocco Isn't it a cyclone object?

Ahh, indeed! I forgot to mention that
The goal is to be pure vanilla, so I'll replace it with:

[*~ -1]
|
[+~ 1]

Thanks @FFW and @alexandros

i wonder how you came up with the filter topology and settings and how you were deciding/balancing the filter curve precision vs the LR runtime difference/phase shift.

HRTF FIRs are a beast, and in my experience it already sounds terrible when you take it over to an fft filter.

@Obineg sorry for the delay!

Actually I'm not the one that designed the filters. I just copied the coefficients from a research published by IRCAM! That is mentioned at the readme file.

Regarding the LR phase difference, that's also a parameter provided by IRCAM (ITD), which I implemented using 2 delay~ lines, one for each channel.

Have you tested out? What do you think of the performance? Does it sound good for your ears?

Thank you for sharing this.
That's nice to have a full Vanilla binaural spatializer!

After a short comparison with the online demo at https://ircam-rnd.github.io/binauralModeled/examples/, I find that the pd implementation reacts a bit differently.
An azimuth of -90 in the Ircam demo put the sound "far away" from the ear.
The implementation result is less obvious, like closer, with less feeling of space.
(Tested with Azimuth: -90 Elevation: 0 in both case, in 44.1, with the breakbeat.wav example in double mono, pd 53.1)

Another difference: with an elevation of 90 (max), we can't perceive any difference when the azimuth is changed.
I guess that's logical, but in the original you can still ear a lot of difference when changing the azimuth.

About the performances, what about a "bake" button that would find the datapoint and switch off the calculations?

In any case thank you for this, that's very helpful already!

@andresbrocco

i just read the starting post again - and as it seems i did not do that before asking that question.

it would have been my next question why you do not use coefficients which somebody else made already.

no i havent tested those yet, but i might come back one day with proposals.