Transcript The single-best-filter model predicts broadband

Spatial release from masking
of chirp trains in a simulated
anechoic environment
Norbert Kopčo
Hearing Research Center
Boston University
Technical University
Košice, Slovakia
UConn 7/23/2003
1
Studies of binaural and spatial hearing
Distance perception in reverberant environments
- is consistent experience necessary for accurate distance perception?
- also, studies looking at other parameters (mono- vs. binaural, anechoic vs.
reverberant, real vs. simulated environments)
“Room learning” and its effect on localization
- is localization accuracy and “room learning” affected by changes in listener position
in a room?
Spatial cuing and localization
- how does automatic attention, strategic attention, and room acoustics affect perceived
location of a sound preceded by an informative cuing sound?
Spatial release from masking
- effect of signal and masker location on detectability/intelligibility of pure tones,
broadband non-speech stimuli, and speech in anechoic and reverberant environments
UConn 7/23/2003
2
Spatial release from masking
of chirp trains in a simulated
anechoic environment
Collaborators
Barbara Shinn-Cunningham (BU) – Thesis Advisor
Courtney Lane (Mass. Eye and Ear Infirmary)
Bertrand Delgutte (Mass. Eye and Ear Infirmary)
UConn 7/23/2003
3
Intro: Spatial release from masking
N
S
N
S
"Spatial unmasking" (or SRM) is an improvement in signal detection threshold
when signal and noise are spatially separated.
UConn 7/23/2003
4
Intro: Spatial release from masking
N
S
N
S
"Spatial unmasking" (or SRM) is an improvement in signal detection threshold
when signal and noise are spatially separated.
Spatial unmasking of low-frequency pure-tone stimuli depends on
- acoustic factors (change in the signal-to-noise energy ratio, SNR, due to
change in location)
- binaural processing (improvement in signal detectability due to signal and
noise interaural cues)
UConn 7/23/2003
5
Intro: Spatial release from masking
N
S
N
S
Spatial unmasking of broadband stimuli depends on
(Gilkey and Good, 1995):
- energetic factors for all stimuli
- additional binaural factors for low-frequency stimuli
UConn 7/23/2003
6
Broadband stimuli: two possible mechanisms
1. auditory system integrates information across multiple channels
2. auditory system chooses single best channel with most favorable SNR
("single-best-filter" model)
Best channel hypothesis supported by comparison of single-unit thresholds
from cat's inferior colliculus to human behavioral data (Lane et al., 2003).
UConn 7/23/2003
7
Current study
Test the single-best-filter hypothesis of
spatial unmasking for broadband and
lowpass stimuli
- measure spatial unmasking for broadband
and lowpass chirp-train signals in noise in
human
- compare performance to single-best-filter
predictions
UConn 7/23/2003
8
Experimental methods: procedure
-
-
-
five listeners with normal hearing
simulated anechoic environment
(i.e., under headphones)
measure detection threshold for
combinations of signal (S) and
noise (N) locations (at 1 m)
signal location fixed at one of
three azimuths (0, 30, 90°)
noise azimuth varies
3-down-1-up adaptive procedure
(tracking 79.4% correct) varying
N level
three-interval, two-alternative
forced choice task
UConn 7/23/2003
9
Experimental methods: stimuli
-
signal: 200-ms 40-Hz
chirp-train
broadband: 0.3 - 12 kHz
lowpass: 0.3 - 1.5 kHz
-
noise: 250-ms white noise
broadband: 0.2 - 14 kHz
lowpass: 0.2 – 2 kHz
-
convolved with nonindividual anechoic human
HRTFs to simulate source
locations
UConn 7/23/2003
10
Single-best-filter model
Filterbank: 60 log-spaced gammatone filters per ear (Johannesma, 1972)
SNR computed in each filter
Single best filter found across all 120 filters
Predicted threshold = -SNR - T0 (T0 is a model parameter fitted to data)
Magnitude
a) Frequency spectra of sample stimuli
Signal at 0°
Noise at 90°
b) Filter with most favorable SNR is chosen
Best channel
Frequency
UConn 7/23/2003
11
Results: broadband stimuli
Data
- spatial unmasking of
nearly 30 dB
Single-best-filter model
- produces accurate
predictions (within 4 dB)
- tends to overestimate
spatial unmasking
- single best filter has high
frequency, so ...
- binaural processing
unlikely to contribute
The single-best-filter model
predicts broadband data
UConn 7/23/2003
12
Results: lowpass stimuli
Data
- thresholds worse than
broadband
- spatial unmasking less
than broadband
Single-best-filter model
- produces accurate
predictions (within 3 dB)
- generally underestimates
unmasking
- underestimation may be
due to binaural
processing
The single-best-filter model
predicts lowpass data
UConn 7/23/2003
13
Results: broadband vs. lowpass stimuli
Data
- for all azimuths,
broadband
thresholds better
than lowpass
Single-best-filter
model
- predicts roughly
equal thresholds
for broadband and
lowpass when
near each other
The single-best-filter model cannot predict lowpass
and broadband data at the same time
UConn 7/23/2003
14
Results: narrowband vs. other stimuli
Data
- thresholds improve with
increasing bandwidth
- highpass and broadband
thresholds similar
- 10 ERB thresholds
approach broadband
- single ERB thresholds
- 10 dB worse than broadband
- approximately equal, indicating roughly equal SNR and information
in each ERB
Single-best-filter model predicts approximately equal thresholds for all
conditions
The single-best-filter model fails to predict thresholds' bandwidth
dependence
UConn 7/23/2003
15
Conclusions: data
For these broadband stimuli, spatial unmasking
- improves thresholds by nearly 30 dB
- is dominated by energetic effects in the high
frequencies
For these lowpass stimuli, spatial unmasking
- improves thresholds by at most 12 dB
- is dominated by low-frequency energetic effects
Binaural contribution is fairly small
Detection thresholds improve with bandwidth
UConn 7/23/2003
16
Conclusions: model
The single-best-filter model predicts the amount of
spatial unmasking for broadband or lowpass
stimuli.
However, the model threshold parameter must differ
in order to achieve these fits.
More generally, the model cannot predict the
observed dependence on signal bandwidth.
UConn 7/23/2003
17
Discussion
It is unlikely that any single-best-filter SNRbased model (regardless of exact
implementation) can account for these
results.
For broadband signal detection in noise, there
appears to be across-frequency integration.
Only a model that integrates information across
multiple frequency channels is likely to be
able to account for these observations.
Brain centers higher than the midbrain seem
necessary for the integration of information
across frequency.
UConn 7/23/2003
18
Acknowledgements
Research supported by AFOSR and National Academy
of Sciences
Steve Colburn and other people in the BU Hearing
Research Center for comments on this works
UConn 7/23/2003
19