Genetic Algorithms: Potential Tools in Fitting Hearing Aids

Nazanin Nooraei, MS., FAAA

Research Audiologist

Principal Investigator: Deniz Bakent

deniz_baskent@starkey.com
2008 Starkey Laboratories, Inc. All Rights Reserved

Motivation
Hearing aid programming can be complicated with many different parameters to adjust: Gain Noise Reduction Feedback Cancellation Directionality Expansion And more

An optimization algorithm such as Genetic Algorithm (GA) might help more efficiently to find the optimal set of parameters for individual users.

What is a Genetic Algorithm (GA)?

The GA is an optimization tool based on principles of evolution. The GA can examine vast number of parameters simultaneously. The GA can optimize the parameters in an efficient manner.

500 1k 2k 4k 6k 0

3: 13 11 30 13 10 81 1: 40 10 30 2: 45 50 81 1: 40 10 32 3: 40 11 30 13 11 81 3: 13 10 32 2: 45 50 30

Genetic Algorithms

dB HL

20 40 80 100

2dB 10dB 40dB

compute fitness (user input)

1: 13 10 30

NO

2: 21 11 81
3: 26 11 56

generate initial population randomly

1: 40 11 81
2: 45 50 32 3: 13 10 30

create new genes from most fit genes (cross-over and mutation)

check convergence criterion

YES

SOLUTION

update gene population with the new genes

Research Study
The GA was used to find the optimal parameters of a new and complex Noise Reduction (NR) algorithm. Optimal parameters had been determined previously by exhaustive listening (many hours of listening).

Parameters of an advanced noise reduction algorithm

1) Offset 2) Slope 3) Time Constant
Total of 6600 possible solutions

Offset: 30 linear steps TC: 11 log space Slope: 20 linear steps

Offset

Research Questions
1) Can the GA produce solutions that are acceptable to the subjects? 2) Are the GA solutions similar to the solutions determined previously by exhaustive listening? 3) Would the subjects like the GA solutions as well as the solutions reported in the previous study? 4) Is the GA a more efficient method in optimizing a new noise reduction algorithm compared with the exhaustive listening approach? 5) Is the GA fast enough to be used as a clinical tool?

Methods
Subjects 13 subjects with normal hearing. Average age of 25
years. Native English Speakers. Nearly half of subjects participated in the previous study (exhaustive listening).

Methods (Cont)
Stimuli
One HINT sentence presented over headphones at 75 dB SPL. The sentence was presented in babble background noise at SNR of 5 dB.

Procedure
3 sets of NR parameters: Slope, Offset, and Time Constant 5-6 paired comparisons in each iteration (10 iterations). GA Search Space: Offset: 30 linear steps Total of 6600 possible TC: 11 log space solutions Slope: 20 linear steps

 7-point scale A/B comparison between two genes. Subjects were asked to choose the sentence that was preferred (quality and comfort)

First
Compared average GA solutions with the best settings from the previous study (Exhaustive Listening)
1) Slope: 95% accuracy 2) Offset: 86% accuracy 3) TC: 100% accuracy

Second
Final Validation
7-point scale paired comparison between the following settings: Original signal, i.e., no NR. NR parameter settings identified previously (Exhaustive Listening). NR parameter settings produced by the GA.

Preference Averaged Across Subjects

Exhaustive Listening GA Settings

No NR

Number of Winners

Finally

Compared the GA average running time with Exhaustive Listening.

Average Running Time GA Study 18 minutes

Minimum

Maximum

10 minutes

34 minutes

Exhaustive Listening

81 hours*

65 hours*

97 hours*

* This is an estimated time and not actually measured

Conclusions
Can the GA produce solutions that are acceptable to the subjects? Yes, the current study showed that the GA can produce solutions that are acceptable to the subjects.

Are the GA solutions similar to the solutions determined previously by exhaustive listening? Yes. The settings produced by the GA were generally close to the best solutions produced previously.

Conclusions (Cont)
Would the subjects like the GA solutions as well as the solutions reported in the previous study? Yes. Subjects generally preferred the GA solutions as well as the solutions reported previously. Is the GA a more efficient method in optimizing a new noise reduction algorithm compared with the exhaustive listening approach? Yes. It is definitely a faster alternative to empirical methods.

Conclusions (Cont)
Is the GA fast enough to be used as a clinical tool? Presently, the GA is not fast enough to be applied in clinical settings. However, it serves as a very valuable tool for research purposes. If the GA is applicable in clinical settings it can be utilized:

in optimizing advanced features of hearing aids. in fine-tuning hearing aid parameters especially in difficult to fit hearing-impaired populations.

Thank you!