Op#miza#on of Loudness-­‐Restoring Hearing Aid Fi9ngs
1
2
2
1
2
1
Kelly Fitz , Shu-­‐Hsien Chu , Mingyi Hong , Mar#n McKinney , Zhi-­‐Quan Luo , Tao Zhang
1Starkey Hearing Technologies, Signal Processing Research Dept., 2University of Minnesota, Dept. of Electrical and Computer Engineering
Mo#va#on
Es#ma#ng Spread of Excita#on
Loudness restoration principle
Excitation should be local
• Restore the specific loudness profile (the loudness density as a function of frequency).
• Specific loudness is estimated by a perceptual model.
Spread of excitation
• Intense low-frequency stimuli are “heard” and contribute to loudness at high frequencies.
0.9
0.8
60
Stimulus unamplified
50
Excess
loudness
without
amplification
0.6
0.5
0.4
Loudness NH
0.3
40
0.2
30 Threshold (dB HL) 45
20
100
50
500
55
1k
Freq (Hz)
65
2k
50
4k
0
100
8k
500
1k
Freq(Hz)
2k
4k
1k
Freq(Hz)
2k
4k
8k
• Apply frequency distance weighting.
• Accumulate weighted loudness loss across
frequency.
• Weighted loudness loss is a measure of
spread of excitation.
25
Weight
20
15
Loudness
10
5
Average spread across channels
8k
Solutions with greater total spread are
penalized (less optimal).
[1]! B. C. J. Moore, “Use of a loudness model for hearing aid fitting. IV. Fitting hearing aids with multi-channel compression so as to
restore ‘normal’ loudness for speech at different levels.,” Br. J. of Audiology, vol. 34, no. 3, pp. 165–177, Jun. 2000.
[2]! B. C. J. Moore, “Dead regions in the cochlea: Diagnosis, perceptual consequences, and implications for the fitting of hearing aids,”
Trends in amplification, vol. 5, no. 1, pp. 1–34, 2001.
Weighted
loudness
0
500
1k
Freq(Hz)
2k
4k
8k
Penalize spread of excitation
• Reduce gain to lessen the
influence on loudness at
distant frequencies.
• May compromise loudness
restoration.
Spread of excitation in each channel
30
25
15
More spread
Less spread
20
15
10
5
0
0
Spread of excitation in each channel
Loudness unamplified
0.1
500
Weighted loudness due to gain in channel 2
Excitation spread
(arbitrary units)
Loudness(Sone)
Response (dB)
Loudness amplified
0.7
70
Loudness with no gain at 500 Hz
0
• Equal to the loudness accrued by adding
gain to the channel.
• Greater penalty for loudness loss at more
distant frequencies.
1
Stimulus amplified
0.2
5
10
Compressor Channel
15
Penalize large gain
discontinuities
• Clinically undesirable
solution not exposed by
the loudness model.
• May compromise
loudness restoration.
Optimized gain at compression threshold
45
10
0
0
35
30
25
20
100
Mean
5
5
10
Compressor Channel
More smooth
Less smooth
40
Gain (dB)
quality[2].
Specific Loudness
80
0.4
Excitation spread
(arbitrary units)
• Can cause excess loudness and loudness growth disturbing artifacts, poor sound
90
• Estimate the loudness lost by removing gain
in each channel.
Loudness with gain applied
0.6
• Average the loudness loss across all stimuli.
• Excessive response of an impaired cochlea to off-frequency stimulation.
Response (1/3rd octave)
Estimate spread in each channel
0.8
Loudness Loss ("Phons"/ERB)
• This is the foundation of the Cambridge Loudness Restoration algorithm
(CAMREST)[1].
Gain applied in a compressor channel should
not influence loudness at distant frequencies.
1
Specific Loudness(Sones/ERB)
• Prescribe gain and compression intended to restore the loudness perceived by a hearingimpaired listener to the loudness that a normal-hearing listener would perceive.
Specific loudness with and w/out gain in channel 2
500
1k
2k
Freq (Hz)
4k
8k
15
Op#miza#on
Balance competing objectives
and feasibility constraints.
• Optimal solution depends on the
relative weight applied to the
various objectives.
Competing objectives
• Restore specific loudness.
• Minimize spread of excitation and excess loudness growth.
• Maintain overall feasibility of the fitting.
Initial Point
• Optimize over all stimuli at once
starting from this solution
Optimality is determined by a model of perceived loudness
Apply non-linear, constrained optimization[3] to find a configuration of gain and
compression parameters that balances competing objectives over a stimulus set.
Restoration-only solution
• Optimal restoration of
loudness for all stimuli
Multi-stage approach
• Solve linear gain-only optimization problem for individual stimuli.
• Construct least-squares compressive approximation to the individual gain-only solutions.
• Optimize compressive solution over all stimuli, using least squares fit as the initial point.
[3] “CONDOR, a new parallel, constrained extension of Powell's UOBYQA algorithm: experimental results and comparison with the DFO algorithm,” vol. 181, no. 1, Sep. 2005.
Response (1/3rd octave)
Specific Loudness
1
Amplified
Unamplified
90
Gain-only problem
• Optimize gain one
stimulus at a time
• Minimize specific
loudness difference
from normal
Loudness(Sones/ERB)
Response (dB SPL)
80
70
60
50
40
Inner/Outer hair cell loss
• Tune the perceptual model
to the individual.
Normal
Unamplified
Amplified
0.8
0.6
0.4
0.2
30
0
20
100
500
1k
Freq (Hz)
2k
4k
8k
100
500
1k
Freq(Hz)
Response (1/3rd octave)
8k
1
Loudness(Sones/ERB)
80
Response (dB SPL)
4k
Specific Loudness
Amplified
Unamplified
90
2k
70
60
50
40
Normal
Unamplified
Amplified
0.8
0.6
0.4
0
500
1k
Freq (Hz)
Response (1/3rd octave)
1
Loudness(Sones/ERB)
80
Response (dB SPL)
4k
8k
100
Specific Loudness
Amplified
Unamplified
90
2k
70
60
50
40
Normal
Unamplified
Amplified
0.8
2k
4k
8k
500
1k
Freq(Hz)
2k
4k
8k
• The quality of a solution is determined by the response of a perceptual model to a set of
treated stimuli.
• The perceptual model is nonlinear, non-invertible, and non-differentiable.
• Parameters of the solution are not mutually independent.
• The optimal solution represents a balance among competing objectives.
0.4
0.2
0
500
1k
Freq (Hz)
Systematic approach to a challenging class of problems in hearing aid
fitting and configuration.
0.6
30
20
100
The Big Picture
0.2
30
20
100
Least-squares fit
• Gain and compression that best
approximate the gain-only solutions
100
500
1k
Freq(Hz)
2k
4k
8k