The Complexities of Understanding
Speech in Background Noise
Stuart Rosen
UCL Speech, Hearing and Phonetic Sciences
First International Conference on
Cognitive Hearing Science for Communication
A caveat about ‘cognition’
• Important aspects of this problem
are not ‘cognitive’ but …
• Cognitive processing …
– relies on adequate sensory
representations, and …
– can compensate for impoverished
sensory representations.
Why is this interesting?
• Most speech is not heard in quiet,
anechoic conditions.
• People vary a lot in how well they
can understand speech in the
presence of other sounds.
– Effects of hearing impairment
– Effects of age
– Auditory processing disorder (APD)?
Some determinants of
performance: I
• The nature of the target speech
material
– Predictability
• context
• number of alternative utterances
• frequency of usage
• size of lexical ‘neighbourhoods’
Some determinants of
performance: II
• The configuration of the environment
– Open air or in a room?
– How ‘dry’ is a room?
• effects of reverberation
– spatial separation between target and
noise
• or, the transmission system (e.g.
mobile telephone)
– distortion & noise added by the system
Some determinants of
performance: III
• Talker characteristics
– Different talkers vary considerably in
intrinsic intelligibility
– Talkers vary their own speech
depending upon demands of the
situation
• hyper/hypo distinction of Lindblom (1990)
– Match between talker and listener
accents
Some determinants of
performance: IV
• Listener characteristics
– Linguistic development
• vocabulary knowledge
• ability to use context
• the presence of language impairments
• L1 vs L2
– Hearing sensitivity and any hearing
prosthesis used
– Neuro-developmental disorders
• Language impairment
• Autism spectrum disorder
• APD
Some determinants of
performance: V
• The nature of the background noises
– level (SNR)
– fluctuations in level
– spectral characteristics
– genuine ‘noise’: aperiodic or periodic?
– and/or other talkers
• how many there are
• speaking your own language or a language you
don’t know
– How ‘attention-grabbing’ the background
noises are
The simplest case:
A steady-state background noise
Much is understood about what
makes one steady noise more or
less interfering than another
spectral shape
SNR
‘Energetic’ masking
• Noises interfere with speech to the
extent that have energy in the same
frequency regions
• Can be quantified in the ‘articulation
index’
• Reflects direct interaction of masker
and speech in the cochlea, which
acts as a frequency analyser.
But noises are typically not
steady …
Fluctuating maskers afford
‘glimpses’ of the target signal
target
glimpses
masker
• People with normal
hearing can listen in
the ‘dips’ of an
amplitude modulated
masker
SRT for VCVs in simple on/off
fluctuations as a function of the duration
of the fluctuation.
Howard-Jones & Rosen (1993) Acustica
 better performance
‘dip listening’ or ‘glimpsing’
 better performance
‘Dips’ can be limited in frequency
(‘checkerboard noise’)
SRT for VCVs in 10 Hz modulations with
different numbers of channels.
Howard-Jones & Rosen (1993) JASA
But maskers can be periodic
too, most importantly, when
speech is in the background.
Miller (1947)
‘The masking of speech’
It has been said that the best place to
hide a leaf is in the forest, and
presumably the best place to hide a
voice is among other voices.
Listening to speech in ‘noise’
Children’s Coordinate Response Measure
Bouncy
in quiet
in steady noise
in modulated noise
against another talker
A useful distinction
• Energetic masking
– maskers interfere with speech to the extent
that have energy in the same time/frequency
regions
– primarily reflecting direct interaction of masker
and speech in the cochlea
– relevance of glimpsing/dip listening
• Temporal and/or spectral ‘dips’ in the masker allow
‘glimpses’ of target speech
• Informational masking
everything else!
Informational masking
• Something to do with target/masker
similarity?
– signal and masker ‘are both audible but the
listener is unable to disentangle the elements
of the target speech from a similar-sounding
distracter’ (Brungart, 2001)
Informational masking: a finer
distinction (Shin-Cunningham, 2008)
• Problems in ‘object formation’
– Related to auditory scene analysis
– similarities in auditory properties make segregation
difficult
• voice pitch, timbre, rate
1 woman, 1 man
• Problems in ‘object selection’
2 men
– Related to attention and distraction
– the masker may distract attention from the target
• e.g., more interference from a known as
opposed to a foreign language
EM & IM appear to operate at different
parts in the auditory pathway
• Energetic masking at the periphery, in the
cochlea
– Early developing abilities
– Increased EM from hearing impairment
– Unlikely to be a factor in APD
• Informational masking at higher centres
–
–
–
–
Late developing abilities?
Increased IM in older listeners?
Increased IM in developmental disorders?
But aspects of IM can be made difficult by
peripheral factors
• e.g., CI users difficulties with auditory scene analysis
little glimpsing for CI users
Nelson et al. (2003)
speech-spectrum-shaped masking noise squarewave modulated added to IEEE sentences
better performance →
normal listeners
CI users
better performance →
not only poor frequency selectivity, but lack of
sensation of voice pitch (poor perception of TFS)
makes auditory scene analysis difficult:
How do you tell the noise from the speech?
But IM can be excessive in the
presence of normal hearing …
 better performance
Children find it hard to ignore
another talker
 better performance
Slow development of abilities
that minimise IM
 better performance
Increased IM in Specific
Language Impairment (SLI)
9 SLI & 10 TD
children aged
6-10 years
CCRM sentences
MSc work of
Csaba RedeyNagy
steady noise
ed
speech
Increased IM in some people with
High Functioning Autism (HFA)
control
better HFA
worse HFA
CCRM sentences in various backgrounds
PhD work of Katharine Mair
• evidence for
a temporal
processing
deficit but …
• not the
crucial factor
in excessive
masking for
speech
Increased IM in some people with
High Functioning Autism (HFA)
control
better HFA
worse HFA
HFA poor
performers (and
younger children)
are highly
susceptible to
informational
masking … but
what aspect?
ASA?
attention?
linguistic aspects?
CCRM sentences in various backgrounds
PhD work of Katharine Mair
An ecologically valid test bed
for evaluating the roles of EM
and IM:
Speech in n-talker babble for
n=1,2,3…∞ talkers
Miller (1947)
Increasing the number of talkers in the
masker
better performance →
SNR (dB)
+12
+6
0
-6
-12
-18
‘It is relatively easy
for a listener to
distinguish between
two voices, but as
the number of rival
voices is increased
the desired speech
is lost in the general
jabber.’
• target words from
multiple males
• babble: equal
numbers of m/f
(1 VOICE is
male)
better performance →
IEEE sentences in n-talker
babble
• What happens as
n increases?
– glimpsing
opportunities 
so EM 
– linguistic
content  so IM
 (selection?)
– number of Fo
contours  so
IM  (ASA)
1-talker voice pitch source with envelopes derived
from n-talker babble
1-talker
2-talker
16-talker
babble-modulated 1-talker F0 (plus with an unmodulated envelope)
2-talker voice pitch source with envelopes derived
from n-talker babble
1-talker
2-talker
16-talker
babble-modulated 2-talker F0 (plus with an unmodulated envelope)
 better performance
Unintelligible maskers on noisevocoded IEEE sentences
noise
Periodicity in the maskers
leads to better performance,
probably through better ASA
It’s easier to ignore a single
F0 contour, rather than two
2 Fo contours
but ...
1 Fo contour
Why improved performance
for steady-state vs 16-talker
envelopes?
Worse still, why glimpsing in
noise?!
Final remarks
• The balance of EM & IM effects presumably varies
with the age and hearing status of the listener
• The linguistic effects seen may represent a
separate aspect of IM apart from object formation
and selection.
• Unraveling the contributions of various factors in
understanding the masking of speech by other
sounds is very important …
– But very complex!
Tack så mycket!
Work supported by:
UCL Speech, Hearing and Phonetic Sciences
National Institutes of Health DC006014
Bloedel Hearing Research Center
Thanks to my collaborators:
Sophie Scott, Katharine Mair, Tim Green, Csaba
Redey-Nagy, Jude Barwell, Zoe Lyall & Arooj
Majeed of UCL
Pam Souza, Northwestern U