auditory model for the speech audiogram - phon.ucl.ac.uk · -physical cues, syntax, semantics,...
TRANSCRIPT
Auditory model for the speech audiogramfrom audibility to intelligibility for words
(work in progress)
Johannes Lyzenga1
Koenraad S. Rhebergen2
1 VUmc, Amsterdam
2 AMC, Amsterdam
Introduction
- History:
- Standard model for sentence intelligibility: SII
- Modified model for sentence intelligibility: SIIcmp
- Comparison of SII and SIIcmp for SRTn and SRTq
- Relation Audibility vs. Intelligibility for sentences
- ? Relationship Audibility and Intelligibility for words ?
- Database of speech audiograms: word scores
- Audibility from modified model for words in quiet
- Relationship Audibility and Intelligibility for words
- Discussion
Speech Intelligibility Index: SII
Assumptions:
- Speech dynamic range of 30 dB, RMS in the middle
- Intensity Importance Function: linear from –15 to +15 dB
RMS
30 dB Dynamic range
Frequency
15 dB “Effective” speech peaks
Level
- SNR calculations executed in frequency bands
- Only the proportion of speech (orange) above the
noise and absolute threshold contributes to the SII
- So: it is basically an Audibility measure!
Calculation of the SII
Frequency
Lev
el (
dB
) 30 dB
Absolute threshold Noise level
Audible speech (orange)
Novel SI model with compression
Introducing compression in the SI model:
- (1) At normal speech levels (ca 65 dB SPL), hearing
in NH listeners is highly compressive
- (2) At very low levels, and for HI listeners, it is not
The SII was designed for NH at normal speech levels (1)
We introduced compression in the calculations (1),
as function of presentation level and hearing loss (2)
And we tried various speech-dynamic ranges
The compression function
Lev
el (
dB
)
After Oxenham, 1995 (PhD thesis)
Schematic diagram of the model(Rhebergen, Lyzenga, Dreschler & Festen, in press)
SI model with compression
Fixed filter:
free field
to eardrum
Fixed filter:
middle ear
Spectrum to
excitation
pattern
Excitation to
specific loudness
(incl compression)
Compress
excitation
pattern
Compressed
excitation pattern
to SIIcmp
FFT-based
Stimulus
Spectrum
Audibility is calculated from the excitation differences for noise and speech: still an Audibility measure
Standard SII predictions
Data set of factory workers:
- Maintenance
work shop for
aircrafts.
- 323 NH: blue
- 65 NIHL: green
- 14 HI: gray
- SIIs in quiet
decrease with
hearing loss !quiet 35dBA 50dBA 65dBA 80dBA
NH
NIHL
HI
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
SII
-∞ 35 50 65 80
Level (dBA)
ANSI S3.5, 1997
-∞∞∞∞ 35 50 65 80
Level (dBA)
SI predictions with compression (1)
quiet 35dBA 50dBA 65dBA 80dBAquiet 35dBA 50dBA 65dBA 80dBA
0,00
0,10
0,20
0,30
0,40
0,50
0,60
0,70
quiet 35dBA 50dBA 65dBA 80dBA
NH
NIHL
HI
Range: 30 dB Range: 35 dB Range: 40 dB0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
SIIcm
p
-∞ 35 50 65 80
Level (dBA)
-∞ 35 50 65 80
Level (dBA)
-∞ 35 50 65 80
Level (dBA)
SI predictions with compression (2)
quiet 35dBA 50dBA 65dBA 80dBA
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
quiet 35dBA 50dBA 65dBA 80dBA
NH
NIHL
HI
quiet 35dBA 50dBA 65dBA 80dBA
Range: 45 dB Range: 50 dB Range: 55 dB0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
SIIcm
p
-∞ 35 50 65 80
Level (dBA)
-∞ 35 50 65 80
Level (dBA)
-∞ 35 50 65 80
Level (dBA)
PTA
-10 0 10 20 30 40 50 60 70
SII
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
SRTq spread of the SII and SIIcmp values
PTA
-10 0 10 20 30 40 50 60 70
SII
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
ANSI
SIIcmp 45dB
SII
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Pe
rcen
t S
en
ten
ce C
orr
ect
(%)
0
10
20
30
40
50
60
70
80
90
100
Relationship Audibility and Intelligibility
SRTs: short, meaningful, sentences in stat. noise
45-dB speech dynamic range optimal for SIIcmp
50% sentences correct gives an SII of appr. 0.22
SRT
Relationship Audibility and Intelligibility
First for words?
Why words…
- Few data sets of psychometric functions for sentences
- From sentence audibility to intelligibility: very complex
- Physical cues, syntax, semantics, prosody, grammar, etc
- From word audibility to intelligibility: less complex
- A lot of data available for words as function of level
- Database: years of clinical measurements at the AMC
- Both speech and pure-tone audiograms available
- Diverse population: NH, M-HI, S-HI, and intermediates
Available data set
Speech audiogram: word scores for at least 3 levels
Pure-tone audiogram: normal audiometric frequencies
Data from 4 years of clinical measurements
NH: 1479
- Age range [18 – 80(!)]
- Avg: 51, SD: 15 years
Not used today:
M-HI: 1967
S-HI: 1314
Inter: 128210
210
310
4−80
−70
−60
−50
−40
−30
−20
−10
0
dB
HL
Hz
average audiogram nrml, interm, imprd, sevr
Results for 30-dB speech dynamic range
Intelligibility and Audibility for Presentation Level
0 10 20 30 40 50 600
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Bosman words clinic: P(c) (bl) and Audibility (rd)
Level (dB SPL)
P(c
) &
Au
dib
ility
Intelligibility vs. Audibility: 30-dB dyn.
50% Intelligibility for Audibility of approximately 0.65
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Bosman words clinic: intelligibility
Audibility
P(c
)
Results for 45-dB speech dynamic range
Intelligibility and Audibility for Presentation Level
0 10 20 30 40 50 600
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Bosman words clinic: P(c) (bl) and Audibility (rd)
Level (dB SPL)
P(c
) &
Au
dib
ility
Intelligibility vs. Audibility: 45-dB dyn.
50% Intelligibility for Audibility of approximately 0.35
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Bosman words clinic: intelligibility
Audibility
P(c
)
Bosman: Intelligibility vs. Audibility in NH
Dyn Range Intelligibility Audibility
30 dB 50% ~0.65
45 dB 50% ~0.35
Sentences 50% ~0.23
Thesis Bosman for NH listeners:
Stimuli Intelligibility Level
Sentences 50% ~20.5
Words 50% ~27.5
Bosman: Word Level for 50% correct is a bit higher ����
Word Audibility needs to be a bit higher: 45 dB Dyn. R.
Discussion
Relationship Audibility and Intelligibility for words:
- Model: plausible relationships for 45-dB speech dyn. range
- The data set shows somewhat different relations than the
data from the thesis of A. Bosman (not shown):
- Refinements needed:
- Separate age groups for NH
- Speech dynamic ranges
- Look at relationship: Sentence Audibility and Intelligibility
Future:
- Maybe we can unearth Intensity Importance functions
- Aim: predict word scores from the audiogram: clinic
Fin
End
Ende
Einde