[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: HRTF failure
We are also using generic HRTFs to virtually arrange sounds in space.
Before begining our experiments, we checked that the "generic HRTF"
way was acceptable by running some preliminary experiments like the
one you described. The principal difference is that we only used
sounds in the frontal hemispace (between -90 (=270°) & 90 ° of Azimuth.
The results were quite acurate but we noted a systematic tendency to :
- overestimate the excentricity for sounds Az of in [-50°;50°] (a
sound presented at a virtual azimuth of 30° is perceived at say 40°)
and an unde
- underestimate excentricity for extreme positions ([-70;-90] &
[70;90]). eg : a -85° az is perceived as "-75°".
These effects are often uncommented but present in figures plotted in
different published papers (in general in papers with a small number
of trials and without intensive training).
Our participant where quite precise (in some experiments we get a
standard deviation <3° in pointing tasks).
You used sound in the back hemispace. I don't know if it is an
important requirement for you but this is known to produce front/back
confusions (some are visible in your plot I think).
To reduce front-back confusion you might consider (see Begault et
- virtually install your participants in a "real room" with some
reverberation that help solving F/B confusions (simulating at least
1st order reverberation)
- You can also use real-time head movement tracking (which imply that
you have a motion tracking system).
- Maybe you can more "compatible" HRTF for each of your subjects (you
have to run a prior experiment in order to choose the best HRTF from
several possibilities instead of only using MIT kemar ones).
I never had accurate pointing response in task involving front & back
sounds in static situations (without head movement tracking).
I hope that this can help and I'm curious about other advises on this
Neuropsychology & Auditory Cognition Team
Le 14 nov. 08 à 11:43, Christian Kaernbach a écrit :
We encounter a problem when trying to place a sound at a virtual
position in space by means of head related transfer functions (HRTF).
We use sounds from the IAPS database (International Affective
Digitized Sounds System, Bradley & Lang) as well as simple white
noise of six seconds duration. We use the Kemar HRTF, the "compact
data" with zero elevation. We convolve the original sound data with
the HRTF data as suggested in the documentation. The final sounds
are presented using Beyer Dynamic DT770 headphones.
We have tested the precision with which our sounds are placed in
virtual space, by presenting them to eight listeners. The listeners
had a touchscreen lying on their lap, with a circle plotted on it,
and they could indicated the direction where they perceived that the
sound came from. We presented to them in total 144 sounds, 72 noises
and 72 IAPS sounds, coming from 36 virtual directions (0°, 10°,
20°...) in randomized order.
The results are shown in a figure that I put in the internet:
The red dots are from IAPS sounds, the yellow dots are from the
noises. The x-axis shows the "true" (virtual) angle, the y-axis
shows the estimated angle. As can be seen in this figure, listeners
could well discriminate between sounds from the left and sounds from
the right. But not more than that. There is a certain reduction of
variance for sounds coming from 90° and from 270°, but there is no
correlation with angle within one hemifield.
Now we are eager to learn from you: What could be the cause for this
A) HRTFs are not better than that.
B) The headphones are inadequate.
C) It must be a programming error (we don't think so)
We are grateful for any help in interpreting the possible cause for
Thank you very much in advance,
Christian-Albrechts-Universität zu Kiel