Temporal-based non-linear hearing aid prescription using a genetic algorithmNorthwestern University, Department of Communication Sciences and Disorders
Andrew T. Sabin, Holly Wiles, and Pamela Souza
Determining the hearing aid parameter settings that optimally compensate for a patient’s hearing loss is critical for successful amplification. As hearing aid signal processing has become more complex, the number of potential parameter combinations has become intractably large. In the face of this complexity, hearing aid parameter settings are most often determined by applying prescriptive formulae to the patient’s audiogram. In nearly all cases, these prescriptive formulae are designed to optimize a value that is derived from the long-term average spectrum of speech (such as SII [e.g.,1] or loudness [e.g., 2]). While these prescriptions take into account the spectral variations in speech, they largely neglect the temporal variations. Several lines of research indicate that accurate perception of temporal variations, especially the temporal envelope, might be particularly important for speech perception in individuals with hearing loss. With this in mind, we began to explore the development of a non-linear hearing aid prescription that takes into account both the temporal and the spectral variations of speech. We used an optimization procedure known as a genetic algorithm.
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
1
2
3
.25 .5 1 2 4 80
20
40
60
80
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
1
2
3
.25 .5 1 2 4 80
20
40
60
80
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
1
2
3
.25 .5 1 2 4 80
20
40
60
80
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
1
2
3
.25 .5 1 2 4 80
20
40
60
80
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
50
100
.25 .5 1 2 4 80
1
2
3
.25 .5 1 2 4 80
20
40
60
80
250 500 1000 2000 4000 8000-60-40-20
0
-1
0
1
-1
0
1
0
50
100
0
50
100
01020
-1
0
1
10 20 30 40Time (sec)
.25 .5 1 2 4 8
-50
0
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 8
-50
0
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 8
-50
0
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 8
-50
0
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 8
-50
0
.25 .5 1 2 4 80
10
20
30
40
50
.25 .5 1 2 4 80
10
20
30
40
50
250 500 1000 2000 4000 8000-60-40-20
0
The input signal
6-Channel Bandpass Fitler
Extract enveloppe
Apply Time Contstants
The Bandpass Filtered Signal
Compute Gain Control Signal
Apply Gain Control Signal
Re-Filter
+
Output
The Multiband Compressor
Gai
n (d
B)
Gai
n (d
B)
Am
plitu
deA
mpl
itude
Am
plitu
dedB
SP
LdB
SP
L
With
in-C
hann
el P
roce
ssin
g
compressorknee
Acr
oss-
Cha
nnel
Pro
cess
ing
Acr
oss-
Cha
nnel
Pro
cess
ing
Introduction The Genetic Algorithm
Parameter Min Max DistributionAttack 0.001 0.3 logRelease 0.001 0.3 logKnee 10 dB SPL 90 dB SPL uniformRatio 1 10 uniformGain 0 dB 70 dB uniform
The Unmodified Channel-Specific
Threshold
UCL
Notch-ClippingAt Threshold
Peak-ReflectionAt UCL
fitness (r) = 0.74
Leve
l (d
B)
Leve
l (d
B)
Leve
l (d
B)
Unmodified
Winner: First Generation
Winner: Final Generation
4. Convergerence:Over the course of generations the best gene becomes more and more “fit.”
5. Stopping:The algorithm was stopped when the same gene was the most fit for three consecutive generations. That gene was taken as the winner.
1. InitializationThie first generation had 100 genes, and each gene had 5 random values
2. Fitness Measurement (all genes)The fitness measure was designed to reflect how well the shape of the temporal envelope was preserved after compression while placing the envelope between the detection threshold and the uncomfortable listeneing level (UCL).
Envelope Post Compressor
Modified Envelope Post Compressor
Compute Fitness(correlation coefficient (r) between modified and unmodified envelopes)
Overview:A genetic algorithm (GA) is an optimization procedure that mimics the mechanism of natural selection. Ideally, this procedure gradually converges on the optimal “gene.” Here each “gene” was an array of values corresponding to compressor settings. The GA was designed to find the “gene” that best preserved the shape of the original temporal envelope. Optimization was conducted separately for each channel.
Test Signal: 21 sec excerpt from the ISTS recording [3] played at both 50 and 80 dB SPL.
3. Making the Next GenerationAfter fitness is evaluated for all 100 genes, a new set of 100 genes is created by these 3 methods.
Preservation:The 10 genes with the highest fitness were included unmodified in the next generation
Mating:45 new genes were created by randomly combining values from the the preserved genes.
Mutation:45 new genes were created by multiplying the values in a preserved gene by a random value drawn from a normal distrubtuion with mean 1 and std 0.25.
Old Gene New Gene
Old Genes New Gene
Old Gene New Gene
Audiogram Attack Release Knee Ratio Gain GA NAL-NL2 [4]Target Gains
Simulation: The GA presciption for 5 hypothetical audiograms
The following rule explains most of the variance in GA target gains. This rule is very similar to DSL (i/o) [5]1. Attack and Release: Very short
set to medians values: 6.0 and 10.7 ms respectively2. Knee: Just above the normal-hearing detection threhsold
set to median value = 17.4 dB SPL3. Ratio: Proportional to the ratio of the dynamic range of the speech signal to that of an individual’s heaing.
RatioF = 0.5063 x (RangeF, speech / RangeF, hearing ) + 0.88684. Gain: Proportional to the magnitude of the indivdual’s hearing loss
GainF = 0.6032 x LossF + 14.6591
Conclusions: The GA created a fit that compressed nearly the entire dynamic range of the speech signal into a the dynamic range of hearing for a given individual at a given frequency.
GA target gains were well predicted by a simple rule (similar to DSL (i/o) [5]) and had far more gain than NAL-NL2 [4], especially in the low frequencies.
The GA-prescribed time constants were very fast. However, the compressor knee was very low, so the compressor behavior was primarily determined by the attack time constant.
How the GA-prescribed fit influences perception and sound quality is currently unknown. Behavioral experiments evaluating these issues are underway.
This work demonstrates that Genetic Algorithms can be used to determine the multi-band compressor settings that optimally preserve temporal envelope shape. However in future work, other GAs can be designed to optimize different fitness functions that incorporate more complex aspects of hearing such as the SII [6] or the non-linear growth of loundess [7].
References: [1] Byrne, D., Dillon, H., Ching, T., Katsch, R., and Keidser, G. (2001). NAL-NL1 procedure for fitting nonlinear hearing aids: characteristics and comparisons with other procedures, J Am Acad Audiol 12, 37-51.[2] Moore, B. C., Glasberg, B. R., and Stone, M. A. Development of a new method for deriving initial fittings for hearing aids with multi-channel compression: CAMEQ2-HF, Int J Audiol 49, 216-227.[3] Inga Holube, Short description of the International Speech Test Signal (ISTS)[4] The National Acoustic Laboratories NAL-NL2 prescription (unpublished)(Cornelisse et al. 1995; Moore and Glasberg 2004)[5] Cornelisse, L. E., Seewald, R. C., and Jamieson, D. G. (1995). The input/output formula: a theoretical approach to the fitting of personal amplification devices, J Acoust Soc Am 97, 1854-1864.[6] American National Standards Institute. (1993).American National Standards Methods for the Calculation of the Speech Intelligibility Index. [7] Moore, B. C., and Glasberg, B. R. (2004). A revised model of loudness perception applied to cochlear hearing loss, Hear Res 188, 70-88.
20 40 60
20
40
60
GA Target Gain (dB)
Rul
e Ta
rget
Gai
n (d
B)
r2 = 0.98
Thre
shol
d (d
B H
L)Th
resh
old
(dB
HL)
Thre
shol
d (d
B H
L)Th
resh
old
(dB
HL)
Thre
shol
d (d
B H
L)
Atta
ck T
Ime
(ms)
Atta
ck T
Ime
(ms)
Atta
ck T
Ime
(ms)
Atta
ck T
Ime
(ms)
Atta
ck T
Ime
(ms)
Rel
ease
TIm
e (m
s)R
elea
se T
Ime
(ms)
Rel
ease
TIm
e (m
s)R
elea
se T
Ime
(ms)
Rel
ease
TIm
e (m
s)
Kne
e (d
B S
PL)
Kne
e (d
B S
PL)
Kne
e (d
B S
PL)
Kne
e (d
B S
PL)
Kne
e (d
B S
PL)
Rat
ioR
atio
Rat
ioR
atio
Rat
io
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Gai
n (d
B)
Frequency (kHz) Frequency (kHz) Frequency (kHz) Frequency (kHz) Frequency (kHz) Frequency (kHz) Frequency (kHz) Frequency (kHz)
Time (sec)
Leve
l (d
B)
Leve
l (d
B)
Leve
l (d
B)
Modified Envelope (dB)
Mod
ified
Env
elop
e (d
B)
Time
Time
Time
Time
Time
Time
Frequency (Hz)
Frequency (Hz)
Gai
n (d
B)
Acknowledgments: The authors thank Harvey Dillon and Scott Brewer at the National Acoustic Laboratories for the use of their NAL-NL2 fitting software.
This work was supported by the Northwestern University Doctor of Audiology program and National Institute on Deafness and Other Communication Disorders grant F31DC009549 (to A. S.) and by R01 DC 0060014 (to PS)