There are 3 messages totalling 295 lines in this issue.
Topics of the day:
1. Experiments with large N (2)
2. low-latency audio I/O for Windows: a report
----------------------------------------------------------------------
Date: Mon, 3 Dec 2007 13:42:14 -0500
From: Robert Zatorre <robert.zatorre@xxxxxxxxx>
Subject: Re: Experiments with large N
Huge samples are very nice if you can get 'em, though such is not
always
the case, alas.
So one thing that I would like to see from people who do have
gigantic N
is to do some analyses to determine at what point the data reach some
asymptote. In other words, if you've collected 1,000,000 people, at
what
earlier point in your sampling could you have stopped, and come to the
identical conclusions with valid statistics?
Obviously, the answer to this question will be different for different
types of studies with different types of variance and so forth. But
having the large N allows one to perform this calculation, so that
next
time one does a similar study, one could reasonably stop after
reaching
a smaller and more manageable sample size.
Has anybody already done this for those large samples that were
recently
discussed? It would be really helpful for those who cannot always
collect such samples.
Best
Robert
-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Robert J. Zatorre, Ph.D.
Montreal Neurological Institute
3801 University St.
Montreal, QC Canada H3A 2B4
phone: 1-514-398-8903
fax: 1-514-398-1338
e-mail: robert.zatorre@xxxxxxxxx
web site: www.zlab.mcgill.ca
Malcolm Slaney wrote:
This music paper has 380k subjects :-)
http://cobweb.ecn.purdue.edu/~malcolm/yahoo/Slaney2007
(SimilarityByUserRatingISMIR).pdf
While Ben Marlin collected another 30k subjects for this
music-recommendation study.
http://cobweb.ecn.purdue.edu/~malcolm/yahoo/Marlin2007
(UserBiasUncertainty).pdf
The underlying data for both papers is available for academic
researchers (fully anonymized, both by song and by user). Send me
email
if you want more information.
- Malcolm
On Dec 1, 2007, at 5:43 PM, Matt Wright wrote:
Trevor Cox recently published the results of an online experiment
about listeners' ratings of sound files on a six-point scale ("not
horrible", "bad", "really bad", "awful", "really awful", and
"horrible"). To date he has 130,000 subjects (!) and about 1.5
million data points:
http://www.sea-acustica.es/WEB_ICA_07/fchrs/papers/ppa-09-003.pdf
Here's the website for his experiment: http://www.sound101.org
Clearly this is related to the "effect of visual stimuli on the
horribleness of awful sounds" that Kelly Fitz pointed out.
-Matt
On Jun 29, 2007, at 12:32 AM, Massimo Grassi wrote:
So far it looks that the experiment with the largest N (513!) is
"The
role of contrasting temporal amplitude patterns in the
perception of
speech" Healy and Warren JASA but I didn't check yet the
methodology
to see whether is a between or a within subject design.
------------------------------
Date: Mon, 3 Dec 2007 13:58:55 -0500
From: "J. Devin McAuley" <mcauley@xxxxxxxxxxxxxx>
Subject: Re: Experiments with large N
This issue nicely highlights the need to report effect size
measures. With a
large enough sample, even the smallest of effects will show up as
reliable!
:)
Best regards,
Devin
-----Original Message-----
From: AUDITORY - Research in Auditory Perception
[mailto:AUDITORY@xxxxxxxxxxxxxxx] On Behalf Of Robert Zatorre
Sent: Monday, December 03, 2007 1:42 PM
To: AUDITORY@xxxxxxxxxxxxxxx
Subject: Re: [AUDITORY] Experiments with large N
Huge samples are very nice if you can get 'em, though such is not
always
the case, alas.
So one thing that I would like to see from people who do have
gigantic N
is to do some analyses to determine at what point the data reach some
asymptote. In other words, if you've collected 1,000,000 people,
at what
earlier point in your sampling could you have stopped, and come to
the
identical conclusions with valid statistics?
Obviously, the answer to this question will be different for
different
types of studies with different types of variance and so forth. But
having the large N allows one to perform this calculation, so that
next
time one does a similar study, one could reasonably stop after
reaching
a smaller and more manageable sample size.
Has anybody already done this for those large samples that were
recently
discussed? It would be really helpful for those who cannot always
collect such samples.
Best
Robert
-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Robert J. Zatorre, Ph.D.
Montreal Neurological Institute
3801 University St.
Montreal, QC Canada H3A 2B4
phone: 1-514-398-8903
fax: 1-514-398-1338
e-mail: robert.zatorre@xxxxxxxxx
web site: www.zlab.mcgill.ca
Malcolm Slaney wrote:
This music paper has 380k subjects :-)
http://cobweb.ecn.purdue.edu/~malcolm/yahoo/Slaney2007
(SimilarityByUserRat
ingISMIR).pdf
While Ben Marlin collected another 30k subjects for this
music-recommendation study.
http://cobweb.ecn.purdue.edu/~malcolm/yahoo/Marlin2007
(UserBiasUncertainty
).pdf
The underlying data for both papers is available for academic
researchers (fully anonymized, both by song and by user). Send
me email
if you want more information.
- Malcolm
On Dec 1, 2007, at 5:43 PM, Matt Wright wrote:
Trevor Cox recently published the results of an online experiment
about listeners' ratings of sound files on a six-point scale ("not
horrible", "bad", "really bad", "awful", "really awful", and
"horrible"). To date he has 130,000 subjects (!) and about 1.5
million data points:
http://www.sea-acustica.es/WEB_ICA_07/fchrs/papers/ppa-09-003.pdf
Here's the website for his experiment: http://www.sound101.org
Clearly this is related to the "effect of visual stimuli on the
horribleness of awful sounds" that Kelly Fitz pointed out.
-Matt
On Jun 29, 2007, at 12:32 AM, Massimo Grassi wrote:
So far it looks that the experiment with the largest N (513!)
is "The
role of contrasting temporal amplitude patterns in the
perception of
speech" Healy and Warren JASA but I didn't check yet the
methodology
to see whether is a between or a within subject design.
------------------------------
Date: Mon, 3 Dec 2007 14:16:16 -0800
From: "Freed, Dan" <DFreed@xxxxxxx>
Subject: low-latency audio I/O for Windows: a report
Dear Auditory List Members:
In October I posted a request for information about low-latency audio
interface devices for use with Windows. I received many helpful
responses. Over the last two months I've had the opportunity to
acquire
several devices and measure their latencies. Since latency
information
is generally not reported (or is incorrectly reported) in manufacturer
specifications, I'm posting my measurement results here.
By "latency", I mean the total end-to-end delay imposed by the device
and its driver, from analog input to analog output. This doesn't
include any additional delay imposed by the software signal processing
(filter group delay, FFT blocking delay, etc.).
Latency was measured by presenting a pulse train to the analog input,
viewing the analog input and output on a dual-trace oscilloscope,
comparing the traces, and visually estimating the delay. Tests were
performed under Windows XP. The PC was running a simple 1-channel
input-to-output copying program that accesses the device through the
ASIO driver interface. Each device was tested at multiple sampling
rates. At each sampling rate, testing was performed using the
shortest
buffer length that the device supports for that sampling rate, so the
measurements are best-case results.
Results are shown below. Sampling rates are in kHz, latencies are in
ms. Buffer length in samples is shown in parentheses. Note that
latencies under 3 ms are achievable at the higher sampling rates with
some devices.
EDIROL UA-1EX [USB device, $80]
32 kHz: 14.2 ms (96)
44.1 kHz: 11.5 ms (96)
48 kHz: 12.0 ms (112)
M-AUDIO FIREWIRE SOLO [FireWire device, $172]
44.1 kHz: 9.0 ms (64)
48 kHz: 8.2 ms (64)
88.2 kHz: 6.5 ms (64)
96 kHz: 6.1 ms (64)
ECHO AUDIOFIRE 4 [FireWire device, $300]
32 kHz: 8.0 ms (32)
44.1 kHz: 6.0 ms (32)
48 kHz: 5.5 ms (32)
88.2 kHz: 3.6 ms (32)
96 kHz: 3.4 ms (32)
RME FIREFACE 400 [FireWire device, $1000]
[at some sampling rates, 48-sample buffer caused bus errors, so used
64-sample instead]
32 kHz: 10.4 ms (64)
44.1 kHz: 6.6 ms (48)
48 kHz: 6.0 ms (48)
64 kHz: 6.2 ms (64)
88.2 kHz: 4.5 ms (64)
96 kHz: 4.2 ms (64)
128 kHz: 4.2 ms (64)
176.4 kHz: 2.9 ms (64)
192 kHz: 2.5 ms (48)
M-AUDIO DELTA 44 [PCI device, $200]
16 kHz: 16.2 ms (64)
22.05 kHz: 11.7 ms (64)
24 kHz: 10.7 ms (64)
32 kHz: 8.1 ms (64)
44.1 kHz: 5.9 ms (64)
48 kHz: 5.5 ms (64)
88.2 kHz: 3.0 ms (64)
96 kHz: 2.7 ms (64)
ECHO LAYLA 3G [PCI device, $500]
32 kHz: 8.8 ms (64)
44.1 kHz: 6.4 ms (64)
48 kHz: 5.9 ms (64)
64 kHz: 4.2 ms (64)
88.2 kHz: 3.1 ms (64)
96 kHz: 2.8 ms (64)
RME MULTIFACE II + HDSP PCI [PCI device, $1049]
32 kHz: 6.5 ms (32)
44.1 kHz: 4.7 ms (32)
48 kHz: 4.3 ms (32)
64 kHz: 3.8 ms (32)
88.2 kHz: 2.8 ms (32)
96 kHz: 2.5 ms (32)
Caveat: running at high sampling rates with short buffer lengths
increases the risk of dropouts. I did some limited listening tests in
all of the above testing conditions and never heard any dropouts,
but I
offer no guarantees.
I'd be happy to answer any questions about my measurements. I hope
this
information is useful.
Dan Freed
Senior Engineer
Dept. of Human Communication Sciences & Devices
House Ear Institute
2100 W. Third St.
Los Angeles, CA 90057 USA
Phone: +1-213-353-7084
Fax: +1-213-413-0950
Email: dfreed@xxxxxxx
------------------------------
End of AUDITORY Digest - 2 Dec 2007 to 3 Dec 2007 (#2007-275)
*************************************************************