University of Southampton

agreeablesocietyΤεχνίτη Νοημοσύνη και Ρομποτική

29 Οκτ 2013 (πριν από 3 χρόνια και 5 μήνες)

70 εμφανίσεις

Stefan Bleeck,

Institute of Sound and Vibration Research,

Hearing and Balance Centre

University of Southampton


Can sparse coding help to overcome
problems caused by hearing loss?



overview of the hearing process


Examples of sparse algorithms for hearing aids and
cochlear implants


Preliminary results



5/11/2012

bleeck@gmail.com

2

5/11/2012

bleeck@gmail.com

3

5/11/2012

bleeck@gmail.com

4

Important for sound localization, linear => boring

5/11/2012

bleeck@gmail.com

5

Important to explain limits of
hearing, linear => boring

5/11/2012

bleeck@gmail.com

6


Contained within bony labyrinth in
temporal bone


Cochlea does hearing


Semicircular

canals+utricle

does
balance


Same mechanism, nerve, evolution,
similar problems


5/11/2012

bleeck@gmail.com

7

7

frequency

mapping

5/11/2012

bleeck@gmail.com

8

8


Stereocilia

detect vibrations
within cochlea.


Introduce half
-
wave rectification


Nonlinear

bleeck@gmail.com

13 orders of magnitude

10

Power (watts/m
2
)

10
-
12

10
-
2

10
-
4

10
-
6

10
-
8

10
-
10

1

5/11/2012

10

ABSOLUTE
THRESHOLD CURVE

membrane
moves 10
-
13

m

Threshold as function of
Frequency

bleeck@gmail.com

5/11/2012

bleeck@gmail.com

11

11


Amplitude (nonlinear amplification)


Frequencies (combination tones)


compression

Demo: sweeps

12

Distance along
BM (mm)

BM displacement (nm)

damaged
“passive”

healthy
“active”

OHCs
inject
energy in
this region

OHCs provide up to


40 dB amplification

(= factor of 100)

Travelling Wave Envelope on Basilar Membrane

due to Pure Tone Stimulus:


50% of 60 year old, 90% of 80 year old


Hearing aids are not good enough


‘damage’

2.4 Billion per year in EU


Lack of research funding today

5/11/2012

bleeck@gmail.com

13

13

5/11/2012

bleeck@gmail.com

14

14

Electron micrographs of cochlear hair cells.

Left: healthy, right: damaged by noise exposure.

5/11/2012

bleeck@gmail.com

15

15

Hearing impairment



loosing audibility,



Also widening of filter



both results in difficulties to understand language, especially in noise


Listening in noise

0 dB

40 dB

-
15 dB

ASR

Normal

Hearing
Impaired

SNR

Word
recognition

100%

0%

Aided

Un
-
aided

50%

5/11/2012

bleeck@gmail.com

18


5/11/2012

bleeck@gmail.com

19

5/11/2012

bleeck@gmail.com

20

Problem: Hearing loss constitutes a bottle neck: not all
information can get through


Solution: extract less, but important information

-
Extract content based on
Information
not on
Energy


-
Specifically speech related
information

2 Neural representation:
(Transformation)

3
Denoising


(sparsification)

1 periphery model

5/11/2012

bleeck@gmail.com

22

‘Sparse’ algorithms developed in our group

noise

5/11/2012

bleeck@gmail.com

23


Filter bank


Non
-
negative matrix factorization


Matrix Z is factorised into two non
-
negative matrices
W and H (basis vectors (5) and activity over time)


(motivated by the processing in CI and auditory
neurons)


Z here is the ‘envelopegram’ (22 channels, 128
pt
)


Factorization using Euclidean cost function:



Sparseness constrained:


5/11/2012

bleeck@gmail.com

24

g(H)= regularity function

λ= sparsity factor

Iterative algorithm to minimize the cost function by gradient decent:

λ

depends on SNR


because of trade
-
off intelligibility
-

quality


low noise: no sparsification


high noise: lots


Task: fine out how!


Online experiment (restricted by speed of hardware)

Offline experiment (unrestricted)

5/11/2012

bleeck@gmail.com

25


5/11/2012

bleeck@gmail.com

26

For ‘bin’, ‘pin’, ‘din’, ‘tin’

Z

W

H


5/11/2012

bleeck@gmail.com

27

5/11/2012

bleeck@gmail.com

28

On
-
line
experimental set
up:


22 channel filter
bank


16
ms

frames


Gaussian
noise

SNR=5 dB



clean

noisy

denoised

time

frequency

5/11/2012

bleeck@gmail.com

30

Results from CI listeners in
online

experiment

(problems with iteration!)

5/11/2012

bleeck@gmail.com

31

Results from CI listeners in
offline

experiment

results for all participants

Averaged

Best sparsification as function of
snr
:

Conclusions:


Sparse coding can help reduce acoustic information in a useful way


Development still in its infancy, hardware restrictions still relevant


High impact research field with lots of potential funding


Strength of our group: clinical evaluation,


weakness at the moment: lack of signal processing experts





5/11/2012

bleeck@gmail.com

33


Hu, H., Li, G., Chen, L., Sang, J., Wang, S., Lutman, M. E., & Bleeck, S. (2011). Enhanced sparse speech coding
strategy for cochlear implants.
European Signal Processing Conference (EUSIPCO)
.


Hu
, H.,
Taghia
, J., Sang, J.,
Taghia
, J.,
Mohammadiha
, N.,
Azarpour
, M.,
Dokku
, R., et al. (2011). Speech Enhancement
via Combination of Wiener Filter and Blind Source Separation.
International Conference on Intelligent Systems and
Knowledge Engineering.



Sang
, J., Hu, H., Li, G., Lutman, M. E., & Bleeck, S. (2011a). Application of a sparse coding strategy to enhance
speech perception for hearing aid users.
British Society of Audiology Short Papers Meeting
.


Sang, J., Hu, H., Li, G., Lutman, M. E., & Bleeck, S. (2011b). Enhanced Sparse Speech Processing Strategy in Cochlear
Implants.
Conference on implantable Auditory Prostheses (CIAP)
.


Sang
, J., Li, G., Hu, H., Lutman, M. E., & Bleeck, S. (2011a). Supervised Sparse Coding in Cochlear Implants.
Conference on implantable Auditory Prostheses (CIAP)
.


Sang, J., Li, G., Hu, H., Lutman, M. E., & Bleeck, S. (2011b). Supervised Sparse Coding Strategy in Hearing Aids.
Annual Conference of the International Speech Communication Association (INTERSPEECH)
.


Bleeck
, S., Wright, M. C. M., & Winter, I. M. (2012). Speech enhancement inspired by auditory modelling.
International Symposium on Hearing
.


Hu, H.,
Mohammadiha
, N.,
Taghia
, J.,
Leijon
, A., Lutman, M. E., Bleeck, S., & Wang, S. (2012). Sparsity Level in a Non
-
negative Matrix Factorization Based Speech Strategy in Cochlear Implants.
EUSIPCO
.


Li, G, Lutman, M. E., Wang, S., & Bleeck, S. (2012). Relationship between speech recognition in noise and sparseness.
International Journal of Audiology
,
51
(2), 75

82. doi:10.3109/14992027.2011.625984


Sang
, J., Hu, H.,
Zheng
, C., Li, G., Lutman, M. E., & Bleeck, S. (2012). Evaluation of a Sparse Coding Shrinkage
Algorithm in Normal Hearing and Hearing Impaired Listeners.
EUSIPCO

(pp. 1

5
).