Transcript Hearing Aids and Hearing Impairments Part II

Hearing Aids and Hearing
Impairments Part II
Meena Ramani
02/23/05
Discussion Time!
Summarize
Room for improvement
Huge Market
Facts on Hearing Loss
Hearing Aids
Cochlea-IHC and OHC
Presbycusis
Decreased
Audibility
Decreased
Frequency
Resolution
Decreased
Temporal
resolution
Decreased
Dynamic Range
Linear-too much gain
Amplification Techniques
CompressiveLinear
Overshoots/undershoots
CompressiveSingle/MultiBand
Multiband vs Singleband
BTE,ITE,ITC,CIC
OHCs:
Sharpen the traveling wave
Provide an amplification for
soft sounds(40-50 dB SPL)
HL in aging ears
Occurs due to damage to OHCs
1) HA has to provide more gain at HFs.
2) HAs less gain at LFs
Noise Removal
3) Fast acting compression
4) Compressive Amplification
Outline




Temporal Resolution
Frequency Resolution
Noise Reduction Techniques
Conclusion
Temporal Resolution




What is temporal resolution?
What happens to temporal resolution for the
HI?
What does poor temporal resolution result in?
Implications for HA design
What is temporal resolution?
Speech has a lot of temporal information like the presence or absence
of acoustic excitation, the periodicity or aperiodicity of excitation
_______, ________,______

Speech Envelope




Modulation Perception



Slowly varying
Carries Information: Consonants, voicing, phoneme boundaries,
syllable boundaries, stress etc.
Lip reading and speech envelope
Changing the depth of modulation of the envelope
Noise and Reverberations
Gap detection threshold



Psychoacoustic measure
For normals: 2.5ms
Relationship between gap detection thresholds and SRTs in noise
(Consonant recognition requires temporal structures)
Temporal resolution for the HI

Experimental setup for Modulation Perception:





Results :



Poor modulation perception is because of reduced listening bandwidth
Same behavior at SL inputs for normals and HI
Results for Gap detection measures:



TMTF- Temporal Modulation Transfer Function
Sinusoidal modulation of broadband noise
Modulation detection threshold in terms of freq.
Comparison with normals: threshold shift /SL
Normals: GDT reduces as the frequency of the noise bands increases
Same behavior at SL inputs for normals and HI
Signals which are made audible to HI have same temporal
resolution as the normals
Temporal resolution for the HI (contd.)




Difference in loudness levels
between the envelope maxima and
minima is > for impaired ear than
normal
This leads one to assume that
impaired ear will perceive
modulation depth changes better
perceptually/Louder
Circles->equal modulation
strength.
Contradicts the TMTF results?


JND vs Perception
Noise also enhanced<Glasberg>
Temporal resolution for the HI (contd.)
Effect of Compression on Modulation






Use compression to provide
loudness correction.
Fast acting/Syllabic compression
3:1 compression
Modulation depth (dB)=20logm
AM factor: (1+msin(wmt))
Reduces the modulation depth
by~9.5dB
Implications for HA design



Syllabic compression can compensate for
abnormal sensitivity to AM
This compression also improves the
discrimination of envelopes having a
DR>10dB
But reducing the spectral cues causes in low
SNR conditions a low SI.
Frequency Resolution




What is Frequency resolution?
What happens to freq. resolution for the HI?
What does poor freq. resolution result in?
Implications for HA design
What is Frequency resolution?
If
change in spectrum of speech causes some
change in shape of excitation along basilar
membrane
=> change exceeds listeners frequency
resolution
Else
=> frequency resolution was not fine enough
to discriminate the spectral changes
Frequency resolution for HI
Statement: Cochlear damage results in poor freq. resolution
But Auditory filter bandwidth increases with stimulus level…
HOW DO YOU MEASURE FREQ. RESOLUTION?



Experimental setup:
 Need normals and the HI to be at the same sensation level(SL)
 Normals: Add broad band background noise to elevate threshold
Results:
 Freq. resolution measured via tuning curves was worse for HI
 More Upward spread of masking since LF slope of filter is
shallower than for normals.
Conclusion:
Freq. resolution is impaired by both:
 Damaged auditory system
 Necessity to listen to high stimulus
What does poor frequency resolution
result in?

Loose Spectral Cues



Inability to distinguish between vowels



Formant peak information is lost
Smooths the internal spectral contrasts
F1 & F2 frequencies Important cue for vowel ID
Increase in upper spread of masking
CVR-Consonant Vowel Recognition

HI have more problems understanding the speech in
noise when compared to normals
Implications for HA design
Fact: For the HI, we have broader auditory filters

Sharpen spectral contrast




Narrow the BW of spectral peaks
Decrease level of spectral valleys
Not too much success since the broad filters
overwhelm the sharpening technique
Multiband/wideband design:


Reduction in spectral cues
For Multiband correlate AGC in each band
Noise Reduction
HI people have abnormal difficult understanding speech in noise.
Noise Reduction

HI need an SNR of 9dB




Broader auditory filters , reduced suppressions
Upper spread of masking
SNR improvement doesn’t often correlate
with improved SI!
Noise removal algorithms


Single-microphone techniques
Multi-microphone techniques
Single-microphone techniques
General Considerations



Single stream has speech+noise
Need to evaluate continuously which frames have
speech and which have noise
Improvements in SNR do not relate directly to
improvements in SI
 Need to evaluate performance of algorithm using
Listening SI tests
Single-microphone techniques
Frequency specific gain reduction




BILL-Bass increase at Low Levels
 For noise reduction, bass decrease at high levels
 Reduces LFs when the average gain in that region is high
 Theoretically should help since the HI LFs mask the HF
 LF has information about consonant features such as nasality,
voicing etc which was lost
Cook et al in 1996 showed that if noise is LF, then HPF the
speech resulted in significant improvement of SI
Festen et al in 1990 Envelope minima technique, reduce gain per
band so that envelope minima (noise) is closer to hearing
threshold level
Dynamic range based technique: attenuation for noise band
inversely proportional to the measured DR
Single-microphone techniques
Spectral subtraction

Subtract spectral magnitude of the noise
estimate from the short term spectral magnitude
of the signal



Assumes stationary noise
Uses same phase for the final noise reduced signal
SNR improves but SI is same.

Removes noise like cues required for fricatives.
Multiple microphone techniques



What is Array Processing?
Omni directional microphones: 15mm separation between any two
Low frequency roll off of 6dB/octave
Figure 7.15. A typical configuration for a twomicrophone (Mic) directional system. The
delay to the back microphone determines
the angle of the null in the directional
pattern.
Figure 7.16. Two directional patterns typically associated with
hearing aid directional microphones. The angle represents the
direction from which the sound is approaching the listener, with 0
degrees representing directly in front of the listener. The distance
from the origin at a given angle represents the gain applied for
sound from arriving from that direction, ranging here from 0 to
25dB. The patterns are a cardioid (left) and a hypercardioid
(right).
Beamforming
Frequency Dependent
Delay and sum
Frequency Independent
Comparison with noise suppressor
• Noise suppressor (NS) is the standard one used on iDEN phones
Noise Cancellation


Use LMS algorithm
Problem is some
speech is fed back to
reference mic and
gets canceled with
noise.
Figure 7.17. A typical two-microphone noise cancellation
system. Ideally, the primary microphone measures a
mixture of the interfering noise and the target speech, and
the reference microphone measures only a transformation
of the interfering noise.
Conclusion






Parameters selection and fitting is a very difficult
problem
Algorithms can make the sound more audible but
not more intelligible
IHC have been ignored so far but they could have a
role too.
It is difficult to get subjective scores from HI
populations
No objective method can account for the nonlinearities introduced by compression
Wearable HAs are an option for research but are
inconvenient
Array fundamentals
 Speaker tracking is not possible

with single microphone
Multiple microphones facilitate
spatiotemporal filtering
 Setup consists of two

microphones with the first
microphone assumed as origin
Distance of the wavefront from
microphone is
d sin  k
 The direction of source is given
by
sin  k 
Nc
(m  1)dFs