I would offer another viewpoint.
If we look at the processing partly in the FFT domain, the time spent at each frequency bin (some delta F) is longer as the sweep length increases. This time window can be considered as a temporal average, over which the SNR increases as the length increases for a given bin. As I mentioned earlier, the SNR gain of averaging and extended sweep length is identical : SNR of deconvolved RIR for avg of 3x20sec sweeps = 1x60sec sweep.
There really isn’t a paradox here.
-Brian FG Katz
LIMSI-CNRS
De : AUDITORY - Research in Auditory Perception [mailto:AUDITORY@xxxxxxxxxxxxxxx] De la part de Trevor Agus
Envoyé : mardi 28 juillet 2015 10:53
À : AUDITORY@xxxxxxxxxxxxxxx
Objet : Re: Optimal sweep duration for BRIR measurements
Dear list,
There's a conundrum in this discussion that:
(1) increasing the duration of the sine-wave sweep increases the signal-to-noise ratio (which seems intuitively true)
versus
(2) increasing the duration of the sine-wave sweep does not affect the signal-to-noise ratio (as John noted; which is literally true if the "signal" is the sweep, whose level is unaffected by duration).
It's a fun paradox, and I don't want to take away from anyone's pleasure by my stab at a resolution... (Potential spoiler alert.)
Is it that the SNR of interest is the impulse-response-to-noise, only over its relatively short time period? As such, a large amount of the energy of the noise (in a long sweep) would be outside of the IR's expected time period (after the convolution stage that 'extracts' the IR), but all the energy of the sine sweep (and its reverberation) would be kept within this time period. So the SNR is constant (if you consider the full duration) yet increased (if you focus on the time period of the IR). Or is there more to it?
All the best,
Trevor
Brian FG Katz wrote:
Dear John,
As others have pointed out, increasing the length of the sweep
increases your signal-to-noise ratio. For large room acoustics, we
typically use sweeps of 20 to 40 seconds. This of course depends on
the size of your room, its reverberation time, and the power of your
source. The weaker the source, the longer the sweep. You should
basically do a test and see what SNR you get. If you need more, and
your measurement chain is at its limit, the only option is longer
sweep. Longer sweeps will not help much with impulsive interruptions,
while averaging will.With a single sweep, basically that freq-range
during the noise is lost. If you are just measuring from the RIR, this
may not be an issue, as measurement parameters are often in
octave-bands, but this is not true for auralization usage; corrupt
data is corrupt data, though I haven’t gone through a thorough study
of this actual case.
The sweep should definitely be longer than the RT, for room
measurements. Then, don’t forget that you need to record the sweep
length PLUS the RT or more if your SNR is better than 60dB!
We have also compared averaging repeated sweeps vs. longer sweeps.
Avoiding the recent developments in overlapping sweep processing, this
basic repetition approach ‘requires’ the decay of the first sweep to
finish before you launch the second sweep. As such, 3x20 second sweeps
take longer than 1 x 60 second sweep, due to the additional pauses. Of
course, without repetitions, you have no backup in case something goes
wrong, like a door slam or something. So, I tend to use repeated
sweeps and take the best 1.
We do all our processing in MatLab, and have never had an issue
(within the last 20 years) of processing long sweeps in real halls.
As you are considering BRIR, and not anechoic HRTFs, you are subject
to the same conditions. If you want to convolve the BRIR directly, you
will need to ensure that the SNR is sufficiently high that the
noise-floor is not audible as a late reverb part of the BRIR. Some
noise extension methods exist from older studies on basic auralization
and scale model RIR auralizations. I cannot imagine a 2s sweep for
BRIR unless you are measuring an office of other room with <1 sec RT.
I think there is a fault in your correlation-based analysis for
reliability, and there are too many factors to consider in comparing
BRIR of different lengths. First, examine your SNR.
Regarding distortion of the source, this should be an issue for the
processing element (different from burning out you speaker). This is
because one of the strengths of the sweep method (when done correctly)
is that any harmonic distortion components of a higher frequency that
the excitation signal at the time or folded back to BEFORE the direct
sound after deconvolution by the excitation signal (as that frequency
has yet to be generated). We presented a through work on this feature,
extending Farina’s earlier works, to general conditions:
M. Rébillat, R. Hennequin, E. Corteel, and B. Katz, “Identification of
cascade of Hammerstein models for the description of nonlinearities in
vibrating devices,” J. Sound and Vibration, vol. 330, pp. 1018–1038,
2011, (doi:10.1016/j.jsv.2010.09.012).
This method lets you actually extract and analyse the different
harmonic distortions (THD etc.) as well as allowing for the modelling
of non-linear responses.
Cheers,
Brian
--
Brian FG Katz, Ph.D, HDR
Resp. Groupe Audio & Acoustique
LIMSI - CNRS
Rue John von Neumann
Campus Universitaire d'Orsay, Bât 508
91405 Orsay cedex
France
Phone. + 33 (0)1 69 85 80 67 - Fax. + 33 (0)1 69 85 80 88
http://www.limsi.fr <http://www.limsi.fr/> web_group:
http://www.limsi.fr/Scientifique/aa/ web_theme:
http://www.limsi.fr/Scientifique/aa/thmsonesp/
*De :*AUDITORY - Research in Auditory Perception
[mailto:AUDITORY@xxxxxxxxxxxxxxx] *De la part de* Anders Tornvig
Christensen
*Envoyé :* lundi 27 juillet 2015 09:07
*À :* AUDITORY@xxxxxxxxxxxxxxx
*Objet :* Re: Optimal sweep duration for BRIR measurements
Hello John,
The sweep method at a single frequency is an approximation to a
steady-state measurement with a pure tone. Longer sweeps give higher
signal to noise ratio per sweep because it spends more time per
frequency.
Short repeated sweeps (but not shorter than the length of the impulse
response) are good, if time-varying or sudden noise that doesn't
average out is likely to contaminate the measurement.
Sweep duration (and "rate" in general) also matters if the system
(room in your case) is nonlinear, time-variant, or both, but that's
another discussion.
Something is wrong with your implementation if the temporal offset of
the impulse responses you measure depends on the sweep duration. You
should be able to check this by connecting the output of your sound
card directly to its input. Also note that wrongly measured or wrongly
computed impulse responses may be very reproducible in terms of
correlation.
Best,
Anders
PhD student in acoustics
Aalborg University, Denmark
------------------------------------------------------------------------
*From:*AUDITORY - Research in Auditory Perception
[AUDITORY@xxxxxxxxxxxxxxx] on behalf of John Culling
[CullingJ@xxxxxxxxxxxxx]
*Sent:* Friday, July 24, 2015 5:25 PM
*To:* AUDITORY@xxxxxxxxxxxxxxx <mailto:AUDITORY@xxxxxxxxxxxxxxx>
*Subject:* Optimal sweep duration for BRIR measurements
Dear all,
Basic Q…
Does anyone have insight into the optimum sweep duration using
Farina's method
for measuring room impulses responses?
More detailed background…
We are planning to make an extensive series of measurements, and in
preparation have
been testing the method using different sweep durations. One way to
check the method
is to correlate the impulses respones from repeated measurements or
those generated
with different durations. To our surprise short sweeps (1-2 seconds)
appear to give more
reliable results (repeated sweeps correlate, r>0.98) than longer ones.
Comparing sweeps
of different durations is a little trickier, because we find a
temporal offset that reduces
the correlation and can only be partially overcome by using
cross-correlation. Nonetheless,
it is apparent that durations from 1 second upwards correlate well,
while going below one
second leads to reliable IRs, but ones that are inaccurate when
compared with those from
longer sweep durations.
Our surprising conclusion is that ~2s should be fine, but Farina
refers to an ISO standard that
recommends very long sweeps (Farina has an example of 50s) to help
overcome noise.
This seems an unintuitive rationale to us, since longer sweeps should
increase both the
signal energy captured and the noise energy, and the method does not
involve averaging
as far as I understand. Longer durations should help address brief
interupting sounds, but
I am unsure if that it what was the idea. In the presence of
continuous noise, we did not
notice any improvement in the IRs produced by longer sweeps.
The nascent plan is to take >1 short sweep for each measurement and
reject IRs that
that don't correlate well with another.
Any insights/advice appreciated,
John.
Prof. John Culling
School of Psychology, Cardiff University
Tel: +44 (0)29 2087 4556
Yr Athro John Culling
Yr Ysgol Seicoleg, Prifysgol Caerdydd
Ffôn : +44 (0)29 2087 4556
Dear John,
As others have pointed out, increasing the length of the sweep increases your signal-to-noise ratio. For large room acoustics, we typically use sweeps of 20 to 40 seconds. This of course depends on the size of your room, its reverberation time, and the power of your source. The weaker the source, the longer the sweep. You should basically do a test and see what SNR you get. If you need more, and your measurement chain is at its limit, the only option is longer sweep. Longer sweeps will not help much with impulsive interruptions, while averaging will.With a single sweep, basically that freq-range during the noise is lost. If you are just measuring from the RIR, this may not be an issue, as measurement parameters are often in octave-bands, but this is not true for auralization usage; corrupt data is corrupt data, though I haven’t gone through a thorough study of this actual case.
The sweep should definitely be longer than the RT, for room measurements. Then, don’t forget that you need to record the sweep length PLUS the RT or more if your SNR is better than 60dB!
We have also compared averaging repeated sweeps vs. longer sweeps. Avoiding the recent developments in overlapping sweep processing, this basic repetition approach ‘requires’ the decay of the first sweep to finish before you launch the second sweep. As such, 3x20 second sweeps take longer than 1 x 60 second sweep, due to the additional pauses. Of course, without repetitions, you have no backup in case something goes wrong, like a door slam or something. So, I tend to use repeated sweeps and take the best 1.
We do all our processing in MatLab, and have never had an issue (within the last 20 years) of processing long sweeps in real halls.
As you are considering BRIR, and not anechoic HRTFs, you are subject to the same conditions. If you want to convolve the BRIR directly, you will need to ensure that the SNR is sufficiently high that the noise-floor is not audible as a late reverb part of the BRIR. Some noise extension methods exist from older studies on basic auralization and scale model RIR auralizations. I cannot imagine a 2s sweep for BRIR unless you are measuring an office of other room with <1 sec RT. I think there is a fault in your correlation-based analysis for reliability, and there are too many factors to consider in comparing BRIR of different lengths. First, examine your SNR.
Regarding distortion of the source, this should be an issue for the processing element (different from burning out you speaker). This is because one of the strengths of the sweep method (when done correctly) is that any harmonic distortion components of a higher frequency that the excitation signal at the time or folded back to BEFORE the direct sound after deconvolution by the excitation signal (as that frequency has yet to be generated). We presented a through work on this feature, extending Farina’s earlier works, to general conditions:
M. Rébillat, R. Hennequin, E. Corteel, and B. Katz, “Identification of cascade of Hammerstein models for the description of nonlinearities in vibrating devices,” J. Sound and Vibration, vol. 330, pp. 1018–1038, 2011, (doi:10.1016/j.jsv.2010.09.012).
This method lets you actually extract and analyse the different harmonic distortions (THD etc.) as well as allowing for the modelling of non-linear responses.
Cheers,
Brian
--
Brian FG Katz, Ph.D, HDR
Resp. Groupe Audio & Acoustique
LIMSI - CNRS
Rue John von Neumann
Campus Universitaire d'Orsay, Bât 508
91405 Orsay cedex
France
Phone. + 33 (0)1 69 85 80 67 - Fax. + 33 (0)1 69 85 80 88
http://www.limsi.fr web_group: http://www.limsi.fr/Scientifique/aa/ web_theme: http://www.limsi.fr/Scientifique/aa/thmsonesp/
De : AUDITORY - Research in Auditory Perception [mailto:AUDITORY@xxxxxxxxxxxxxxx] De la part de Anders Tornvig Christensen
Envoyé : lundi 27 juillet 2015 09:07
À : AUDITORY@xxxxxxxxxxxxxxx
Objet : Re: Optimal sweep duration for BRIR measurements