Transcript Speech Intelligibility Index

Taiwan, July, 2009
Prescribing hearing aids
and the new NAL-NL2
prescription rule
Harvey Dillon
National Acoustic Laboratories and
The Hearing Cooperative Research
Centre
National Acoustic Laboratories, Sydney, Australia
Acknowledgment
Individual contributions: Gitte Keidser,
Teresa Ching, Matthew Flax, Richard
Katsch, Karolina Smeds, and Justin Zakis
 Data gatherers: Several audiologists at
NAL and from Australian Hearing
 Sponsors: Bernafon, GN ReSound, Oticon
foundation, and Siemens

Using a prescription
- easy
Adult
Child
Measure hearing thresholds
(dB HL)
Measure hearing thresholds
(dB HL or dB SPL)
Measure individual RECD
(or estimate from age)
Enter into manufacturer software
(hearing aid auto adjusted to
approximate prescription)
Enter into manufacturer software
(hearing aid auto adjusted to
approximate prescription)
Verify with real ear
measurement
Adjust amplification to
better match prescription
Adjust hearing aid in coupler via
computer to better match
prescribed coupler gain
Deriving a prescription
- hard
Overall approach to prescription
Psychoacoustics
Assumptions,
rationale
Speech science
Theoretical
predictions
Compare
Empirical
observations
Final formula
National Acoustic Laboratories, Sydney, Australia
NAL-NL1


Maximize speech intelligibility, with overall
loudness at or below normal
Derived from an optimization procedure
combining:
 the
Speech Intelligibility Index formula (modified)
 the 1997 loudness model of Moore and Glasberg

Prescribes gain-frequency responses that
 make
loudness of speech bands approximately
constant across frequency
 at medium levels agree with NAL-RP
National Acoustic Laboratories, Sydney, Australia
Post NAL-NL1

The underlying principles and assumptions
have been evaluated
 Is
the rationale appropriate?
 How can we better predict speech
intelligibility?
 Do new hearing aid users prefer less gain
than experienced hearing aid users?
 What compression is preferred by hearing aid
users with severe to
profound hearing loss?
Maximizing speech
intelligibility
Is this a better rationale than
loudness normalization?
National Acoustic Laboratories, Sydney, Australia
Loudness perception of speech
bands
Loudness
Normal loudness
Equal loudness
100
1000
Frequency in Hz
10000
National Acoustic Laboratories, Sydney, Australia
NAL-NL1 (speech intelligibility maximisation) vs
IHAFF (loudness normalisation)
8
S
teeplyslopingloss
M
ildflatloss
M
od/sevflatloss
6
NAL-NL1
preferred
Diferenceinsatisfactionscore
(NAL-NL1-IHAFF)
4
2
(r = 0.65,
p = 0.001)
0
-2
0
2
4
6
8
10
12
14
rm
sd
iffe
re
n
c
eb
e
tw
e
e
nN
A
L
-N
L
1a
n
dIH
A
F
F
Source: Keidser and Grant 2001
Desired loudness
Feedback suggested that
NAL-NL1 was too loud!
National Acoustic Laboratories, Sydney, Australia
Loudness; adults, medium input level
(N = 187)
National Acoustic Laboratories, Sydney, Australia
Loudness, experience
National Acoustic Laboratories, Sydney, Australia
Gain preference over time
N = 11
Source: Keidser, O’Brien, Yeend, & McLelland (submitted)
National Acoustic Laboratories, Sydney, Australia
Loudness; adults, low and high input levels
Smeds et al. 2006
Zakis et al. 2007
Preferred gain deviation from
NAL-NL1 re gain preferred at 65
dB SPL in dB
1.5
1
0.5
0
-0.5
-1
-1.5
-2
50
Suggest that the
compression ratio should
be slightly higher,
at least for clients with
mild
80 and moderate
hearing loss
Input level in dB SPL
National Acoustic Laboratories, Sydney, Australia
Compression by severe and
profound hearing loss
1.1
1:1
1.0
0.9
0.8
0.7
0.6
1.8:1
0.5
0.4
Average 1/CR in
3:1
0.3
0.2
45
50
55
60
65
70
75
80
Average HTL in LF band (dB HL)
Source: Keidser, Dillon, Dyrlund, Carter, and Hartley (2007)
85
90
National Acoustic Laboratories, Sydney, Australia
Children
Preference: diary rating
20
18
Canada
Australia
16
No. of children
14
12
10
8
6
4
2
0
DSL
Source: Ching 2003
Undecided
NAL
Prescription
DSL
Undecided
NAL
National Acoustic Laboratories, Sydney, Australia
Adults – congenital or acquired?
Preferred gain deviation
from NAL-RP (dB)
0
-2
-4
-6
-8
-10
-12
Congenital
(N=15)
Acquired
(N=28)
-14
LFA
HFA
National Acoustic Laboratories, Sydney, Australia
Loudness; children

An increase of gain
 is
more likely to lead to greater speech intelligibility at
low input levels where speech is most limited by
audibility
 is less likely to cause noise-induced hearing loss for
low input levels than for high input levels

Increase gain for low input levels with a
progressive decrease in increased gain with
increased input level (i.e. higher CR)
National Acoustic Laboratories, Sydney, Australia
Increased compression ratio
Output level
Children
NAL-NL1
Adults
Input level
National Acoustic Laboratories, Sydney, Australia
NAL-NL2:





Keep maximizing speech intelligibility rationale
Change the intelligibility modelling
Prescribe less gain for adults, but more gain for
children
Increase the CR for adults and children with mild
and moderate hearing loss, while restricting the
CR for those with severe/profound hearing loss
Introduce gain adaptation for new hearing aid
users, and a gender effect
National Acoustic Laboratories, Sydney, Australia
Any questions at this point?
Predicting speech
intelligibility
Is our formula for predicting
maximum speech intelligibility
optimized?
1/3 octave SPL
30
Freq
Audibility:
... 5
Importance:
...
... 16
x
x
0.001
0.002
=
=
... 17
x
... 0
x
0.003 0.002
=
=
0.005 ... 0.032 ... 0.051 ... 0
= 0.30
Speech Intelligibility Index
Sum
SII = ∑ Ai Ii
Audibility
Importance
But intelligibility gets worse if we make speech too loud!
Speech intelligibility also depends on …
Level distortion
Normal hearing people perform poorer at
high speech levels
Level distortion factor

1
0
0
73
140
Speech level (dB SPL)
Percent Correct
The transfer function
Sentences
100
80
60
Nonsense syllables
40
20
0
0
0.2
0.4
0.6
0.8
Speech Intelligibility Index (SII)
1
Observed and Predicted performance
1400 - 5600 Hz
100
MS
MF
SF
SS
Percent correct
80
60
40
20
0
12
24
36
12
24
36
12
24
36
12
24
36
Sensation level (dB)
Ching, Dillon & Byrne, 1998
National Acoustic Laboratories, Sydney, Australia
Optimizing speech intelligibility
1
NAL-NL1: Frequency
dependent hearing
loss desensitization in
quiet introduced
 Dead
regions?
(Baer et al., 2002;
Moore, 2004)
 Desensitization is
different in noise?
(Turner & Henry,
2002)
2 kHz
Effective audibility

0 dB HL
0.8
80 dB HL
0.6
0.4
100 dB HL
0.2
120 dB HL
0
0
10
20
30
Sensation level (dB)
Source: Ching, Dillon, Katsch & Byrne (2001)
40
National Acoustic Laboratories, Sydney, Australia
New study


75 test subjects
Measurements
 Hearing
threshold levels
 Outer hair cell function

click-evoked otoacoustic emissions
 Frequency


psychophysical tuning curves
cochlear dead regions – TEN test
 Speech


resolution
perception in quiet and noise
consonants
sentences
Source: Ching et al. 2005
Subjects


20 adults with
normal hearing
55 adults with
sensorineural
hearing loss
-20
0
20
40
60
80
100
Hearing thres
– mild to
profound
– Experienced
hearing aid
users
Audiogram
120
250
500
1000
2000
Frequency (Hz)
4000
8000
Speech perception



Stimuli: Filtered speech
– CUNY sentences
– VCV syllables
Shaping:
– POGO prescription
Conditions:
– Quiet at high and low
sensation levels
– Babble Noise
Headphones:
Sennheiser HD25
Level (dB SPL)

HP7
LP14
HP14
LP28
HP28
LP56
LP7
80
60
40
20
0
250
500
1000
2000
Frequency (Hz)
4000
8000
Audibility and Speech intelligibility – H.I.
VCV
1.0
0.8
0.6
0.4
Proportion correct
0.2
0.0
0.0
0.2
0.4
Calculated SII
0.6
0.8
1.0
LP7: Q
LP7: N
LP14: Q
LP14: N
LP28: Q
LP28: N
LP56: Q
LP56: N
HP28: Q
HP28: N
HP14: Q
HP14: N
HP7: Q
HP7: N
Deficit = Sansii - SIIeff
Percent Correct
100
80
Deficit = 0.6 - 0.4
= 0.2

60
40
20
SIIeff
SIIansi
0
0
0.2
0.4
0.6
0.8
Speech Intelligibility Index (SII)
1
VCV deficit vs CUNY deficit
0.25
0.20
0.15
R=0.77
0.10
0.05
0.00
-0.05
CUNY SIIansi-SIIeff
-0.10
-0.15
-0.20
-0.25
-0.15
-0.10
-0.05
0.00
0.05
0.10
VCV SIIansi-SIIeff
0.15
0.20
0.25
0.30
Intelligibility and audibility
1
m
p
30
Sensation level (dB)
Variation of m with HL
m
1.0
mp
0.5
ms
0
Hearing Threshold (dB HL)
BKB, VCV and CUNY
Optimizer results: 3 data sets
BKB
Q
&
N
VCV
CUNY
Desensitisation for hearing loss
1
Effective audibility
0.9
0.8
0
20
40
60
80
100
120
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
20
40
Sensation level (dB)
60
Should we use anything
other than the audiogram
to predict speech
intelligibility?
Psychoacoustic correlations – 2 kHz
Matrix Plot (Prof and Psy 11 March 09.sta 659v*75c)
HL2k
PTC2k_Q10
C2k
El2k
Cognition
AgeLim
Correlations
Age
PTC
HL
OAE
Cognit
TEN
Multiple regression
including HL causes:
correlations between age and PTC / OAE / TEN to disappear
correlations between cognition and PTC / OAE / TEN to disappear
PTC
Age
HL
OAE
TEN
Cognition
Likely intermediate effects
Cognition
?
Age
Cardiovascular
Mechanical
PTC
Stria
OAE
OHC
TEN
Noise
IHC
HL
Deficit (VCV & CUNY), HL, Q10, OAE, TEN, Cog, & Age: 5600 Hz low pass
es-SIIeff (VCV Avg(QH QL N) LP5600)
SIIdes-SIIeff (CUNY Avg(QH QL N) LP5600)
L56HL
Q10 LP56
COAE L56
EL LP56
Cognition
AgeLim
Implications for prescription
Pure tone thresholds critical
Knowledge of temporal resolution, frequency resolution,
dead regions adds relatively little to prediction of
intelligibility
Age and cognitive ability affect all frequency bands similarly
 no effect on gain needed
Deriving a
prescription
Using the intelligibility and
loudness models
The rationale for NAL proceudres
Maximize calculated speech intelligibility ,
but
Keep total loudness less than or equal to normal
The two key ingredients
1.
2.
A loudness model
An intelligibility model
Calculating loudness
Loudness model of Moore and Glasberg (2004)
Allowance for hearing loss
External &
middle ear
Free field
speech level
Filtering into
auditory
bands
Input to
cochlea
Excitation
level
Calculate
loudness
per band
Loudness
per band
Sum
across
bands
Total
loudness
Deriving optimal gains - step 1
Speech spectrum
& level
Loudness
model
Gain-frequency
response
Normal
loudness
Compare
Intelligibility
achieved
Intelligibility
model
Amplified speech
spectrum
Audiogram
Loudness
model
Loudness
(hearing
impaired)
The audiograms
Rejection criterion :
-30<= G <=60 , where G is the slope
sum(H(f))/3 <=100 , where f is in the set {0.5, 1, 2} kHz
Inverted hearing loss profiles used
The audiograms, continued
Deriving optimal gains - step 1
Audiogram 1
Speech level 1
Optimal gain frequency response
Audiogram 1
Speech level 2
Optimal gain frequency response
Audiogram 1
Speech level 3
Optimal gain frequency response
Audiogram 2
Speech level 1
Optimal gain frequency response
1200 audiograms x 10 speech levels  12,000 gain–frequency
responses, each at 20 frequencies from 125 Hz to 10 kHz
The result of step 1 ……
Gainf
HLf
HL3FA
f
SPL
Deriving optimal gains - step 2
• Fit a multi-dimensional equation to the data
• Gain at frequency f depends on: f, HL at all frequencies, SPL
• Apply constraints:
•
No compression for speech < 50 dB SPL
•
Low compression ratio for profound loss for fast compression
•
No gain at very, very low frequencies (e.g. 50 Hz)
•
No gain at very, very high frequencies (e.g. 20 kHz)
Gain
50
Input SPL
Multi-dimensional equation:
A neural network
H250
H500
H1000
H2000
H8k
G250
G500
G100
G200
G8k
0
0
SPL
In summary:
This is how NAL-NL2 targets
differ from NAL-NL1 targets,
and vary with client profile
Insertion gain (dB)
30
NAL-NL2
25
20
15
10
0
250
Preliminary
5
500
1000
2000
4000
2000
4000
2000
4000
Frequency (Hz)
500
1000
2000
35
4000
1) Female new user
2) Male experienced user with dead region
3) Child user
Insertion gain (dB)
250
30
NAL-NL2
25
20
15
10
5
0
250
35
30
25
35
20
15
10
5
0
250
500
500
1000
Frequency (Hz)
NAL-NL1
Insertion gain (dB)
Insertion gain (dB)
HTL (dB HL)
35
0
10
20
30
40
50
60
70
80
90
100
1000
Frequency (Hz)
2000
4000
30
NAL-NL2
25
20
15
10
5
0
250
500
1000
Frequency (Hz)
“A challenge for the profession
is to devise fitting procedures
that are scientifically defensible
and the challenge for the
individual audiologist is to
choose the best procedures
from whatever are available”
Denis Byrne, 1998
Thank you for
listening
www.nal.gov.au