Transcript powerpoint

Localizing a speech target
in a multitalker mixture
Norbert Kopčo 1, Virginia Best 2, and Simon Carlile 2
1Technical
2School
May 21, 2009
University of Košice, Košice, Slovakia
of Medical Sciences, University of Sydney, Sydney, Australia
ASA 09 Portland
Introduction
Spatial separation of sources enhances speech perception
In complex environments (e.g., with multiple talkers), spatial perception
also important for “sorting” acoustic scene into objects and focusing
attention on sources of interest (Brungart et al 2001; Freyman et al 1999;
Kidd et al 2005; Best et al 2007; Shinn-Cunningham 2008)
Relatively few studies actually measured localization of speech in a
multitalker environment (Yost et al., 1996; Hawley et al.1999; Drullman and
Bronkhorst 2000; Brungart et al. 2006)
2
Experiment and Goals
Study horizontal localization of speech in a multitalker environment
Question 1: How does presence of maskers influence localization
performance?
Evaluate the effect of maskers on biases/variability in responses.
Question 2: Is performance affected by a priori knowledge /
uncertainty about distribution of masker locations?
Compare performance when masker distribution fixed vs. varied from
trial to trial.
3
Setup and masker patterns
Room:
approx. 3 m x 5 m
Speakers:
only presenting targets
presenting targets and
maskers (see panel b)
Pattern 1
Pattern 2
Pattern 3
Pattern 4
Pattern 5
4
Methods
Stimuli:
Target: word “two” spoken by a female talker
Maskers: 4 different monosyllabic words,
spoken by 4 male talkers (all longer than target)
Target-to-Masker energy ratios: 0 dB or -5 dB
Task:
Subjects pointed head to perceived target location
Subjects asked to indicate location only if target heard
(5 catch trials with no target per block to monitor obedience)
Conditions (separate blocks):
- Control: No masker
- Fixed: Masker pattern fixed across block of trials
- Mixed: Masker pattern randomly chosen for each trial
Analysis:
- evaluate the effect of the maskers on the bias and st.dev.
re. control . Plot across-subject means + standard errors
5
Detection
Miss Rate
N=7
25
25
20
20
Percent
Percent (out of 5 catch trials per block)
False Alarm Rate
15
15
10
10
5
5
0
0
0 dB
-5 dB
TMR
Fixed
Mixed
0 dB
-5 dB
TMR
Detection worse at lower TMR, similar in both uncertainty conditions
6
 Rightward
Bias due to Maskers
TMR 0 dB
Fixed
Mixed
20
Mean Response Bias [°]
0
-20
-50
0
50
-50
0
50
-50
0
50
-50
0
50
-50
0
50
-50
0
50
Target Azimuth [°]
-50
0
50
-50
0
50
TMR -5 dB
20
Leftward 
0
-20
-50
0
50
-50
0
50
Compression strongest for targets near peripheral maskers
7
Masker Uncertainty and Bias
Bias due to Masker Uncertainty
Left   mean [°]  Right
Mean Response [°]
Average across TMR
Fixed
Mixed
20
0
-20
-50
0
50
-50
0
50
-50
0
50
-50
0
50
-50
0
50
0
50
-50
0
50
Difference between meanFixed and meanMixed
10
5
0
-5
-10
-50
0
50
-50
0
50
-50
0
50
Target Azimuth [°]
-50
When masker pattern fixed throughout a block, responses biased away from maskers
8
Response Variability
TMR 0 dB
Fixed
Mixed
20
Standard Deviation [°]
10
0
-50
0
50
-50
0
50
-50
0
50
-50
0
50
-50
0
50
-50
0
50
Target Azimuth [°]
-50
0
50
-50
0
50
TMR -5 dB
20
10
0
-50
0
50
-50
0
50
Complex effect of target location, masking pattern, uncertainty and TMR
9
Uncertainty and Response Variability
Difference between Fixed and Mixed conditions
9
6
3
0
1
2
3
Pattern
4
5
Patts 1-3 (grouped maskers):
Averaged x-location, no Mixed-Fixed diff.
If not averaged, fixing pattern
- helps for off-masker targets
- hurts for on-masker targets
TMR 0 dB
TMR -5 dB
Std.Dev MIX - Std.Dev FIX [°]
Standard Deviation [°]
12
TMR 0 dB
TMR -5 dB
FIX
MIX
Masker Uncertainty
Helps 
 Hurts
Average across locations
6
3
0
-3
On- OffMasker
On- OffMasker
On- OffMasker
On- OffMasker
On- OffMasker
1
2
3
Pattern
4
5
10
Uncertainty and Response Variability
Difference between Fixed and Mixed conditions
9
6
3
0
1
2
3
Pattern
4
5
Patts 1-3 (grouped maskers):
Averaged x-location, no Mixed-Fixed diff.
If not averaged, fixing pattern
- helps for off-masker targets
- hurts for on-masker targets
TMR 0 dB
TMR -5 dB
Std.Dev MIX - Std.Dev FIX [°]
Standard Deviation [°]
12
TMR 0 dB
TMR -5 dB
FIX
MIX
Masker Uncertainty
Helps 
 Hurts
Average across locations
6
3
0
-3
On- OffMasker
On- OffMasker
On- OffMasker
On- OffMasker
On- OffMasker
1
2
3
Pattern
4
5
Patts 4-5 (distributed maskers):
Effect of uncertainty independent of loc. or TMR
- patt 4: uncertainty hurts performance
- patt 5: uncertainty helps performance
11
Summary
1. Mixture has complex effects on localization bias and variability
- depending on masker pattern, location of target re. maskers, and TMR
- compression of mean localization responses near peripheral maskers
- increases in standard deviations, in particular when maskers distributed
2. Trial-to-trial randomization in the distribution of speech maskers (i.e.,
masker location uncertainty) modulates the effect of masking:
- sometimes exaggerating it (as expected)
- but sometimes reducing it (unexpected)
These modulatory effects could be due to
- changes in strategy
- adaptation
12