Validation of a binaural model for speech intelligibility in noise and

Validation of a binaural model for speech intelligibility in noise and

Validation of a binaural model for
speech intelligibility in noise and reverberation
Jelfs S., Culling J. F. & Lavandier M.
BRIR Target
BRIR Interferer 1
BRIR Interferer 2
.
.
.
BRIR Interferer N
Predictions in
anechoic*
Noise interferers and frontal target sentences in all
experiments
BMLDs
Better-ear S/N ratios
Binaural S/N ratio
predicting SRT
Better-ear listening
Peissig & Kollmeier (1997)
Hawley et al. (2004)
Culling et al. (2004)
- Subjective adjustment procedure in
German
- Interferers: one (17 azimuths between
0° and 360°), two (17 az.+105°), or three
(17 az.+105°+255°)
- Adaptative procedure in English
- One, two or three interferers at one,
two or three of a range of azimuths (-30°,
0°, 30°, 60° and 90°)
- Binaural and monaural listening
- Same as Hawley et al. (2004) with
three interferers
- Conditions with both cues (FF), no
binaural unmasking (ILD-only) or
no better-ear listening (ITD-only)
Anechoic => head-related impulse responses (HRIRs)
instead of BRIRs for the predictions
6
Predicted SRT (dB)
3
When original HRIRs not available (P&K1997,
Hal.2004, B&P1988) => HRIRs from Gardner and
Martin (1994)
Energy levels
1 interferer
2 interferers
3 interferers
0
-3
-6
* Jelfs et al. (2011), Hear. Res. (in press)
-6
-9
open symbols: monaural, r = 0.99
closed symbols: binaural, r = 0.99
r =0.98
-12
-12
-12
-9
-6
-3
0
3
6
-9
-6
Both cues (FF)
ITD only (dT)
ILD only (dL)
0
-3
-12
-9
1
1 interferer
2 interferers
3 interferers
0
Predicted SRT (dB)
* computationally efficient version of
Lavandier & Culling (2010) model
Interaural parameters
(ΦT - ΦI, ρI)
Intelligibility
Predicted SRT (dB)
BRIR: binaural room impulse response
BMLD: binaural masking level difference
Φ: interaural phase difference
ρ: interaural coherence
S/N ratio: signal-to-noise ratio
SII: speech intelligibility index
SRT: speech reception threshold
Binaural unmasking
Gammatone Filterbank
BRIRs between sources & ears
SII-weighted
integration
Model*
-3
0
-1
-2
-3
-4
-5
r =0.95
-6
-6
Observed SRT (dB)
-5
-4
-3
-2
-1
Observed SRT (dB)
Observed SRT (dB)
Bronkhorst & Plomp (1988)
- Adaptative procedure in Dutch
- One interferer at 0°, 30°, 60°, 90°, 120°, 150°, or 180°
- Conditions with both cues (FF), no binaural unmasking (dL)
or no better-ear listening (dT)
Predicted SRT (dB)
-5
FF (Both cues)
dL (ILD only)
dT (ITD only)
Main discrepancy associated with a HRIR difference at 90°
-10
-15
r =0.88
-20
-20
-15
-10
-5
Observed SRT (dB)
Predictions in
reverberation*
Noise interferers, target sentences and
real-room BRIRs in all experiments
Adaptive procedure in English
(except B&B2006)
Experiment 1
Experiment 2
Beutelmann & Brand (2006)
- One interferer at three azimuths (-25°, 0°
or 25°) and two distances (0.65 or 5 m), and
a target at (25°, 0.65m)
- Meeting room 1
- Conditions with both cues (BRIR) or no
binaural unmasking (SEIR for Spectral
Envelope Impulse Response)
- One interferer in 12 selected conditions
involving three azimuths (-25°, 0° or 25°),
5 distances (0.65, 1.25, 2.5, 5 and 10 m)
and 4 rooms (corridor, L-shaped, lecture
hall, meeting 2)
- Adaptative procedure in German
- One interferer (8 azimuths between
-140° and 180°) and a frontal target
- Two rooms (cafeteria, office)
* Lavandier et al. (2011), J. Acoust.
Soc. Am. (under review)
Experiment 3
- One, two or three interferers in 4 configurations (see below) at two distances (0.65 or 5 m)
- Meeting room 1
- Conditions with both cues (BRIR) or no
binaural unmasking (SEIR)
near (0.65 m)
or far (5 m)
0.65 m
1int.
Conclusions
2 int.
bilateral
3 int.
 model predicting speech reception thresholds in combined noise and reverberation
 accounts for the effects of binaural unmasking, head-shadow and room colouration
 for multiple interfering sources and real-room acoustics
 reproduces a range of data sets from the literature (≠ languages and procedures)
 without parameter fitting
0
1