Precedence-based speech segregation in a virtual auditory environment

1
Precedence-based speech segregation in a virtual
auditory environment

• Brungart, Simpson Freyman (2005)

2
The Precedence Effect

• Sounds produced in areas with multiple surfaces
  give rise to reflections. Many copies of a sound
  reach a listeners ears. The direct sound arrives
  first.
• With complex sounds like speech, early
  reflections tend to perceptually fuse with the
  direct sound (the Haas effect).
• The direct sound dominates localisation the
  precedence effect.

• /- 0.5 ms gt summing localisation

lt- Perceived direction
two sources perceived
D gt 1 ms gt precedence effect
D gt 20 ms gt echo threshold
3
Masking

• the amount of interference one stimulus can
  cause in the perception of another stimulus.
  (Yost and Nielsen, 1977)
• The elevation in threshold of a target signal due
  to the presence of a masker.
• Energetic masking
• masking that results from competition between
  target and masker at the periphery of the
  auditory system, i.e., overlapping excitation
  patterns in the cochlea or auditory nerve (AN).
  (Durlach et al., 2003)
• Informational masking
• Non-energetic masking
• Central masking
• difficulty segregating the audible acoustic
  components of the target speech signal from the
  audible acoustic components of a perceptually
  similar speech masker. (pp. 3241).

4
Some Assumptions

• Speech target
• Random noise masker purely energetic masking?
• Speech masker energetic and informational
  masking?
• So if an experimental manipulation affects the
  amount of masking produced by the speech masker
  but not the noise masker this is due to a
  reduction in informational masking?
• Seems reasonable

5
The Basic Experiment
Freyman et al., 99 free-field. Brungart et al.
virtual auditory space over headphones
F-F Baseline masking
F-R Release from masking regardless of type of
masker
F-RF Release from masking with speech but NOT
with noise masker
6
Experiment 1
Adding delayed copy of noise to front presented
stimulus drops performance to baseline
Adding delayed copy of speech to front hardly
makes any difference
Note using a speech recognition task which is
resistant to energetic masking - Therefore large
informational masking component?
7
Interpretation

• The precedence effect causes the listener to
  localise the RF masker off to the right, which
  helps auditory selective attention attend to the
  target speech, hence reducing informational
  masking.
• This doesnt affect the noise masker because it
  has no informational masking effect adding it
  to the front just increases its energetic masking
  effect.
• BUT The effect is also observed when the delay
  is negative, so that the first copy of the masker
  comes from the front (i.e. F-FR). (Freyman et al.
  1999)
• Precedence should localise the masker to the
  front in this condition so why the release from
  masking with a speech masker?

8
Experiment 2

• What is the effect of varying the delay between
  the two masker presentations between / 64 ms?
• For a noise masker?
• Very little.
• Some release from masking at delays which cause
  notches in the spectrum of the masker far
  enough apart to be resolved by the ear
• For a single-speaker speech masker?
• Little effect of delay, positive or negative,
  until the echo threshold is exceeded
• For a two-speaker speech masker? Much more
  variation, but still substantial release from
  masking. Possibly some release from energetic
  masking effects
• Note that as speakers are added, multi-speaker
  babble approaches speech-shaped noise.

9
A Puzzle

• There is virtually no difference between positive
  and negative delays with the single-speaker
  masker and not much of an advantage with the
  two-speaker masker
• What is going on here?
• Two possibilities (actually 3, but Ill come back
  to this)
• 1) The effect is not based on perceived location,
  but on timbre or source width
• 2) Even when the copy of the masker added to the
  front leads the one from the right, the one to
  the right pulls the perceived location off a
  little so that it is perceived somewhere between
  front and right
• If (2) is the case, then shifting the apparent
  location of the target to match that of the
  masker, should abolish the release from masking

10
Experiment 3
Position of target varied from 0o to 60o In 5o
steps, at 7 different delay values from to
4ms.

• U-shaped performance curves for all 3 maskers at
  D 0 ms. Masker heard midway between front and
  right.
• For the two-speaker masker, when there is a lag
  (ve D) gt 0.5 ms, subjects do best when target is
  located near the front (0o). As expected
• When there is a lead (-ve D) gt 0.5 ms, subjects
  do best when target is located to the right.
• BUT the minimum performance is found around 10o
  NOT at 0o

11
Conclusions

• This would appear to support the hypothesis
  mentioned earlier
• BUT why is there not a similar minima around
  50o when there is a positive delay?
• Also energetic and informational masking do not
  seem to have been completely separated by this
  paradigm as was first thought
• AND no mention is made of the phenomena of the
  BMLD
• Whenever the phase or level differences of the
  target signal at the 2 ears are not the same as
  those of the masker, ability to detect or
  identify the target improves
• Inversion of the signal at one ear gives better
  performance than delaying it so not just
  segregation by spatial separation
• Large BMLDs occur when target and masker are not
  subjectively well separated
• Hearing is sensitive to the profile of interaural
  decorrelation across frequency
• This could potentially explain why negative
  delays are as useful as positive delays adding
  a delayed copy of the masker at the right changes
  the interaural correlation of the masker relative
  to the target
• But this still wouldnt explain the difference
  between speech and noise