"From
ITU-T Workshop on
Speech to Audio: bandwidth extension,
binaural perception"
Lannion, France, 10-12 September 2008
OPTIMUM FREQUENCY RESPONSE
CHARACTERISTICS FOR WIDEBAND
TERMINALS
H. W. Gierlich, F. Kettler, S. Poschen
HEAD acoustics GmbH
A. Raake, S. Spors
Deutsche Telekom Laboratories (Germany)
Lannion, France, 10-12 September 2008
International
Telecommunication
Union
PART 1: Receiving Frequency Response
Characteristics
Receiving frequency response of 3
wideband phones (handset mode, 3.4 ear, 8N, free-field )
L/dB[Pa/V]
30
Tolerance Scheme acc. ETSI ES 202 739
20
10
0
-10
provides good listening
speech quality
-20
Test Conditions
ƒ
Use 3.4 or 3.3 artificial ear
ƒ
Use 8N application force
ƒ
Use artificial voice or
composite source test
signal
ƒ
NEW: use DRP to FF
correction
-30
100
200
500
f/Hz
2000
5000
-> Is the tolerance scheme really desirable?
Lannion, France, 10-12 September 2008
International
Telecommunication
Union
2
New Subjective Experiments: Pretests with
Experts Listeners
Parametric EQ
Player
…
looped playback of
speech signal
individual adjustment of preferred frequency
response characteristics
conducted with different types of phones
* published DAGA 2007, March 07 and ETSI STQ, April 08
Lannion, France, 10-12 September 2008
International
Telecommunication
Union
3
The Results - Some Examples
L/dB[Pa/V]
20
10
0
-10
-20
-30
-40
100
200
500
f/Hz
2000
5000
-> wide range of personal preferences
Lannion, France, 10-12 September 2008
International
Telecommunication
Union
4
Step 2: Rank Ordering of the Response
Characteristics
Phones with different EQ-Settings applied
at HATS
Parametric EQ
Test
System
…
Ranking of the different EQ-Settings by
experts:
Results:
clear “winner” setting for most phones
“favorite” frequency responses similar
for all phones
Lannion, France, 10-12 September 2008
International
Telecommunication
Union
5
Example Result
• “optimum” response characteristics for 3 phones
L/dB[Pa/V]
20
10
0
-10
-20
-30
-40
100
200
500
Lannion, France, 10-12 September 2008
f/Hz
2000
5000
International
Telecommunication
Union
6
Step 3: Listening Only Test
Test Conditions
4 speakers: 2 female, 2
male
4 wideband phones,
each combined with
several frequency
responses
1 artificial bandwidth
extension
1 ortho-telephonic
reference
1m
Lannion, France, 10-12 September 2008
Representation &
Assessment
43 conditions x 4
speakers, each
assessed by 12 naïve
test persons
Assessment of “speech
sound” on a five point
MOS-scale:
5
4
3
2
1
–
–
–
–
–
excellent
good
fair
poor
bad
International
Telecommunication
Union
7
•
•
•
Lannion, France, 10-12 September 2008
Li
n
Te
l2
Fa
v1
N
8N
N
8N
8N
13
Fa
v3
N
2N
13
Fa
v2
Fa
v3
Fa
v3
Fa
v3
Te
l3
Te
l2
Te
l3
Te
l2
Te
l2
N
8N
13
Fa
v1
Fa
v2
Te
l2
Te
l2
13
3N
8N
8N
8N
2N
8N
op
t1
Fa
v2
Fa
v3
Fa
v1
Fa
v1
N
N
2N
2N
t8
N
Fa
v2
Li
n
Te
l1
Te
l1
Te
l7
Te
l7
Te
l3
op
E
8N
op
t8
Fa
v2
FF
N
N
op
t8
Fa
v1
Fa
v3
Li
n
Te
l2
t8
N
AB
op
op
t8
N
2N
13
Fa
v1
Li
n
Te
l2
Te
l3
Te
l7
Te
l3
Te
l2
Te
l1
Te
l1
Te
l3
N
t8
N
op
t8
Fa
v3
Li
n
D
iff
Te
l1
Te
l2
op
3N
8N
2N
8N
t8
Fa
v1
FF
op
t1
Fa
v3
N
8N
13
Fa
v2
op
N
t8
N
Fa
v1
D
iff
Te
l1
Te
l1
op
Fa
v2
FF
Te
l7
Te
l7
Te
l1
Te
l1
Te
l1
Te
l7
Te
l1
N
N
t2
N
t8
op
t2
op
op
Fa
v2
FF
Li
n
FF
D
iff
t8
1
P
R
ef
op
o_
D
iff
Te
l1
Te
l1
Te
l1
Te
l1
Te
l1
rth
t8
N
2_
B
op
R
ef
FF
o_
O
rth
Te
l2
O
Te
l2
Results of LOT
Analysis of answers over sounds (all speakers)
5
Average and confidence interval
4
3
2
1
whole quality range is covered
“winners”: ortho-telephonic reference, flat frequency response with DF or FF EQ
-> extract for new tolerance scheme
International
Telecommunication
Union
8
Conclusion - New Tolerance Scheme for
Receiving
Possible new tolerance scheme with diffuse field reference:
L/dB[Pa/V]
20
10
0
-10
-20
-30
-40
100
200
500
f/Hz
2000
Lannion, France, 10-12 September 2008
5000
9 frequency responses
with MOS-LQS ≥ 3.6
(average over all
speakers)
black lines: “most tight”
tolerance possible
red line: possible
“smoothed” version of
lower tolerance line
Results discussed in
ETSI STQ and ETSI NG
DECT, to be discussed in
ITU
International
Telecommunication
Union
9
PART 2: Frequency Response in
Sending
Sending frequency response of 3 wideband
phones
(handset mode, 3.4 ear, 8N)
L/dB[V/Pa]
20
10
Tolerance Scheme acc. ETSI ES 202
739
Problem:
-20
-30
-40
200
500
f/Hz
2000
Lannion, France, 10-12 September 2008
5000
ƒ All phones provide good listening
speech quality (TMOSwb = 4.2)
0
-10
100
ƒ Tolerance scheme given by ETSI ES
202 739 is mostly meet for all
ƒ Today’s and future phones are more
and more used in noisy
environments
ƒ Transmission of wideband noise
ƒ People start talking with “Lombard
Effect”
International
Telecommunication
10
Union
Desirable Frequency Response Char. in
Noisy Environments
Setup of the subjective experiment (Step 1 & 2):
ƒ Binaural recording of different
realistic background noises
(here: car, pub, café, living room, call
centre) Æ STEP 1
ƒ Speaker in anechoic conditions,
noise playback via closed headphone
1
Æ initiation of Lombard effect
- recording with omni-directional
microphone
- recording synchronously to the
background noise Æ STEP 2
2
- (recording of Lombard speech for
one female and one male
speaker, each for 5 background
noises)
Lannion, France, 10-12 September 2008
International
Telecommunication
11
Union
Desirable Frequency Response Char. in
Noisy Environments
Step 3: Recording of Speech plus Background Noise
1
2
VoIP Terminal
under test
RCV
3
4
Lannion, France, 10-12 September 2008
Synchronous playback of
corresponding Lombard
speech (2) via HATS with
background noise
Transmission of Lombard
speech and noise via 3
wideband phones
(G.722) (3)
Use “optimum” receiving
frequency response to
simulate receiving side
(4)
International
Telecommunication
12
Union
Pretests with Experts Listeners
Equalizer adjustment by 7 expert listeners
for 3 wideband phones with 5 background
noises each
Parametric EQ
…
looped playback
Adjustment of preferred frequency response
to achieve best overall sound quality
(speech and noise)
Lannion, France, 10-12 September 2008
International
Telecommunication
13
Union
Results of Pretests
Expert‘s equalizer settings for …
non-stationary cafeteria noise (68 stationary car noise (69 dB
dB(A))
Cafe, Tel 1
L/dB
(A))
Car, Tel 1
non-stationary call centre noise
L/dB
20
10
10
10
0
0
0
-10
-10
-10
-20
-20
-20
-30
-30
-30
-40
-40
-40
500 700 f/Hz
Cafe, Tel 7
2000 3000 5000 9000
L/dB
200 300
500 700 f/Hz
Car, Tel 7
L/dB
500 700 f/Hz
2000 3000 5000 9000
200 300
500 700 f/Hz
Call Centre, Tel 7
2000 3000 5000 9000
L/dB
20
20
20
10
10
10
0
0
0
-10
-10
-10
-20
-20
-20
-30
-30
-30
-40
-40
-40
-50
-50
100
200 300
Lannion, France, 10-12 September 2008
500 700 f/Hz
2000 3000 5000 9000
phone 7
200 300
100
2000 3000 5000 9000
-50
100
-50
-50
100
phone 1
200 300
L/dB
20
-50
100
Call Centre,
Tel 1
(58
dB(A))
20
100
200 300
500 700 f/Hz
2000 3000 5000 9000
International
Telecommunication
14
Union
Results and Assumptions
Results:
equalizer settings …
show similar characteristics for all experts for the different
noise-telephone combinations
are noise level dependent: high and low passes are inserted
vary only slightly for one type of noise in combination with
different phones
Assumptions:
Depending on the noise level more or less strong high and low
passes are inserted in order to limit the annoyance due to the
background noise – although also the speech sound is moving
towards narrowband speech!
For further improvement of the speech intelligibility an
emphasis around 2…3 kHz is applied.
Lannion, France, 10-12 September 2008
International
Telecommunication
15
Union
Listening Test: Filters Applied
L/dB
30
fullband
20
A
50
100
200
f/Hz
2000
5000
20
10
0
0
0
-10
-10
-10
B
-30
50
L/dB
30
100
200
f/Hz
2000
10
100
200
f/Hz
2000
Depending on the
results of the expert’s
pretest, the following 8
filters were created
and used for the
listening tests
50
100
200
f/Hz
2000
5000
Lannion, France, 10-12 September 2008
f/Hz
2000
2000
-30
5000 10k
L/dB
30
20
10
-10
-10
F
-20
-30
50
10k
L/dB
30
G
200
f/Hz
0
20
10
100
200
0
high-pass rising 2 dB
per octave
50
100
-20
strong high-pass,
medium low-pass Æ
symmetric shape
10
E
-30
5000 10k
50
20
-10
-20
-30
L/dB
30
medium high-pass,
mod. low-pass Æ
symmetric shape
20
C
-20
5000 10k
0
50
20
10
-20
D
L/dB
30
strong high-pass,
super mod. low-pass
10
10k
moderate high-pass,
super mod. low-pass Æ
symmetric shape
L/dB
30
medium high-pass,
super mod. low-pass
5000 10k
100
200
f/Hz
2000
5000
-20
-30
10k
L/dB
30
medium high-pass
with emphasis
around 2 kHz
20
10
0
0
-10
-10
H
-20
-30
50
100
200
f/Hz
2000
-20
-30
5000 10k
International
Telecommunication
16
Union
Listening Test 1
Test conditions – LOT 1
ƒ speech transmitted via 1 wideband
phone (flattest sending frequency
response, “traditional big” handset,
no noise reduction)
ƒ 5 background noises
ƒ filtered with 8 filters extracted from
expert pre-test
ƒ all filtered with optimum frequency
response of receiving frequency
response test
ƒ Due to Lombard effect up to 15 dB
level difference between listening
samples -> loudness adjustment to
provide “same” loudness for all
samples
Presentation & Assessment
ƒ 2 speakers: 1 female, 1 male
Æ36 conditions x 2 speakers, each
assessed by 16 naïve test persons
ƒ Diotic representation with open,
free-field equalized headphones
ƒ Assessment of “overall quality” on a
five point MOS-scale:
5 – excellent
4 – good
3 – fair
2 – poor
1 – bad
RCV
4
Lannion, France, 10-12 September 2008
International
Telecommunication
17
Union
Results of LOT 1
Findings:
The higher the background noise level is the stronger
the Lombard effect is. -> Lower MOS scores
Some results for males better than for females (females
tend to sound shrill and sharp for higher background
noises in conjunction with the Lombard effect)
No clear preference for one response characteristics,
only tendencies
Full bandwidth is preferred for most noises
Strong high-pass characteristics is not preferred
Lannion, France, 10-12 September 2008
International
Telecommunication
18
Union
Listening Test 2
Test conditions – LOT 2
ƒ
Separate assessment for 3 noises (café, car, living room)
ƒ
Reduced number of filters per noise
ƒ
Additional “simulated” conditions with 5/10 dB worse SNR for
each noise and filter
ƒ
Listening level adjusted to 73 dB SPL active speech level
ƒ
Diotic representation with free-field equalized head phones
ƒ
3-fold assessment by naïve persons (similar to P.835):
- listening effort (complete relaxation possible, no effort required –
attention necessary, no appreciable effort required – moderate effort required
– considerable effort required – no meaning understood with any feasible
effort)
- speech sound quality (excellent – good – fair – poor – bad)
- overall quality (excellent – good – fair – poor – bad)
ƒ
Assessment in steps of 0.5 MOS
Lannion, France, 10-12 September 2008
International
Telecommunication
19
Union
Results LOT 2 (café noise).
Café, Listening Eff., female
5,0
Café, Speech Sound, female
Café, Speech Sound, male
5,0
Café, Listening Eff., male
4,5
4,5
4,0
4,0
3,5
3,5
3,0
3,0
2,5
2,5
2,0
2,0
1,5
1,5
1,0
1,0
Fullband
Filter A
SNR-5,
mod HP,
sym
D
5,0
med HP,
SNR-10, SNR-5, str
sym
str. HP, str. HP, sym
TP
E
F
F
str HP,
sym
F
SNR-5,
2dB/oct
HP
G
2dB/oct
HP
4,0
3,5
3,0
2,5
2,0
1,5
1,0
Fullband
SNR-5,
mod HP,
sym
med HP,
sym
SNR-10, SNR-5, str str HP, sym SNR-5, 2dB/oct HP
str. HP, str. HP, sym
2dB/oct HP
TP
Lannion, France, 10-12 September 2008
SNR-5,
mod HP,
sym
med HP, SNR-10, SNR-5, str
sym
str. HP, str. HP, sym
TP
str HP,
sym
SNR-5,
2dB/oct
HP
2dB/oct
HP
G
Café, Overall Q., female
Café, Overall Q., male
4,5
Fullband
Female / male voice:
- For all parameters and both SNRs: only slightly
different MOS scores for full-band, moderately
band-pass and 2 dB / oct. filtered versions
- strong band-pass filtered version leads to
significantly lower MOS scores for the male
voice than for the female voice. Since this
background noise has no dominant low
frequency components, the limited bandwidth
impairs the speech quality instead of reducing
the annoyance due to the background noise.
Furthermore the MOS score of the example with
the lower SNR is about 0.5 MOS higher.
International
Telecommunication
20
Union
Results of LOT 2 (car noise)
Car, Listening Eff., female
5,0
Car, Speech Sound, female
5,0
Car, Listening Eff., male
Car, Speech Sound, male
4,5
4,5
4,0
4,0
3,5
3,5
3,0
3,0
2,5
2,5
2,0
2,0
1,5
1,5
1,0
1,0
Fullband
Filter A
SNR-5,
mod HP,
sym
D
med HP,
sym
E
SNR-10, SNR-5, str
str. HP,
HP, sym
str. TP
F
F
str HP,
sym
F
SNR-5,
2dB/oct
HP
G
2dB/oct
HP
G
4,5
4,0
3,5
3,0
2,5
2,0
1,5
1,0
Fullband
SNR-5,
mod HP,
sym
med HP,
sym
SNR-10, SNR-5, str
str. HP,
HP, sym
str. TP
SNR-5,
mod HP,
sym
med HP,
sym
SNR-10, SNR-5, str
str. HP,
HP, sym
str. TP
str HP,
sym
SNR-5,
2dB/oct
HP
2dB/oct
HP
Male voice:
Car, Overall Q., female
Car, Overall Q., male
5,0
Fullband
str HP,
sym
SNR-5,
2dB/oct
HP
Lannion, France, 10-12 September 2008
2dB/oct
HP
- no significant difference for all parameters between
fullband and filtered versions for original and 5 dB
reduced
SNR;
- moderately band-pass and 2 dB/oct. high-pass
filtered
versions tend to be slightly better than full-band
version
Female voice:
International
Telecommunication
- moderately and strong band-pass filtered
21
Union
versions result
in slightly higher MOS scores than full-band
Sending Frequency Response in Noisy
Environments - Final Conclusions
Proposal for frequency response shape in sending direction under noisy
conditions
ƒ “acoustical noise reduction”
by applying a moderate or –
depending on the noise – even
medium band-pass filter may
be helpful:
L/dB
30
20
10
ƒ acoustical noise reduction
with only slightly affect the
speech sound
ƒ noise reduction algorithms
may be adjusted less
“aggressively”
0
-10
-20
-30
50
100
200
f/Hz
2000
5000
ƒFilter should be applied
adaptively - only if a
background noise is
detected at the terminal’s
microphone
ƒ Investigation should be
repeated for different
wideband noise reduction
algorithms
Lannion, France, 10-12 September 2008
10k
L/dB
30
Recommended
filter still matches
ES 202 739 and
740 tolerance
schemes!
20
10
0
-10
-20
-30
50
100
200
f/Hz
2000
5000
10k
International
Telecommunication
22
Union