IRISA 2003 SPEAKER RECOGNITION SYSTEM

1
IRISA 2003 SPEAKER RECOGNITION SYSTEM

• 1sp DETECTION
• Limited Data
• M. BEN, G. GRAVIER, A. OZEROV F. BIMBOT
• for the ELISA consortium

NIST Speaker Recognition Workshop, June 24-25,
2003
2
Outline

• IRISA 2003 system
• Introduction
• Description
• NIST03 SRE results
• Experiments
• Front-end
• Modeling
• Score normalization
• Conclusions

3
IRISA 2003 system

• Introduction

? IRISA is a member of the ELISA consortium ?
IRISA 2003 system is based on a newly developed
audio segmentation software audioseg Web
links - IRISA/METISS http//www.irisa.fr/meti
ss/accueil.html - ELISA consortium
http//elisa.ddl.ish-lyon.cnrs.fr
4
IRISA 2003 system

• Description
• front-end

• 20 ms frames every 10 ms
• 24 filter bank over 340 - 3400 Hz ? 16 LFCC
• RASTA filtering (secondary system)
• deltas delta log-energy are added
• frame selection bi-gaussian modeling of the
  energy with ML classification of the frames
  (speech/silence)
• global feature normalization (zero mean, unit
  var.)

5
IRISA 2003 system

• Description
• background modeling
• speaker models

• gender-dependent background models
• 256 GMMs with diagonal covariance matrices
• prim. system cellular data (NIST01)
• second. system cellularlandline data
  (NIST01)
• adapted from the background models with MAP
  estimation of the parameters (mean only
  adaptation)

6
IRISA 2003 system

• Description
• scoring

• frame score
• log-likelihood ratio using the 10-best
• matching gaussians in the background model
• utterance score
• NT number of frames in the utterance

7
IRISA 2003 system

• Description
• score normalization DT-norm

• D-norm
• D(spk) symmetric Kullback-Leibler
  distance
• between the speaker (spk) and
• the background models
• DT-norm
• mean and standard deviation of the
• D-norm scores of the test utterance
• using cohort impostor models (50 mal.
• 50 fem. from NIST01 SRE)

8
IRISA 2003 system

• NIST03 SRE results 1sp-limited
• DET curves

• 2 systems submited
• IRI_1 primary
• baseline system
• IRI_2 secondary
• RASTA front-end
• mixed cell.land. data for world models

DCF min actual IRI_1
0.3176 0.3205 IRI_2 0.3333
0.3396
9
Experiments

• Front-end frame selection

• speech/silence classification based on a
• bi-gaussian modeling of the frame energy
• ML classification
• or
• threshold-based selection ?
• ( t ?2 - c.?2 )
• constant coef. to optimise

G1(?1 ,?1)
G2 (?2,?2)
energy
10
Experiments

• Front-end frame selection

• speech/silence classification based on a
• bi-gaussian modeling of the frame log-energy
• ML classification
• or
• threshold-based selection ?
• ( t ?2 - c.?2 )
• constant coef. to optimise

G1(?1 ,?1)
G2 (?2,?2)
log-energy
11
Experiments

• Front-end frame selection

• SYS_fs1 ML selection (E)
• SYS_fs2 optimal threshold-based selection (E)
  c 0.8
• SYS_fs3 ML selection (LogE)
• SYS_fs4 optimal threshold-based selection
  (LogE) c 2.5
• energy (E) bi-gauss. modeling with ML selection
  of the frames performs the best
• drastic selection about 50 of the frames are
  discarded !

NIST 03 SRE data
12
Experiments

• Front-end feature normalization

- st-norm short-term norm. (0 mean, unit
var.) on a sliding window (3 sec.) - lt-norm
long term norm. (0 mean, unit var.) on all
features

• st-norm is applied before frame
• selection
• lt-norm can be applied before or
• after frame selection
• SYS_fn1 lt-norm frame selection
• SYS_fn2 st-norm frame selection
• SYS_fn3 frame selection lt-norm

NIST 02 SRE data (subset)
13
Experiments

• Front-end feature normalization

• - SYS_fn5 frame selection lt-norm
• baseline system (prim.)
• SYS_fn6 st-norm frame selection
  lt-norm
• short-term normalization does not seem to work
  well (buggy?)
• long-term normalization at the end of front-end
  seems to be crucial
• best results obtained with frame selection
  followed by long-term normalization of remaining
  features

NIST 03 SRE data
14
Experiments

• Modeling

• Does size matter ?

• - SYS_nbg1 256 component GMMs
• (baseline)
• SYS_nbg2 2048 component GMMs
• no gain of performance with 2048 gaussians in
  the mixture
• may be due to the frame selection process which
  remove a large amount of frames (?)

NIST 02 SRE data (subset)
15
Experiments

• Score normalization

• SYS_sn1 no score norm.
• SYS_sn2 T-norm
• SYS_sn3 DT-norm
• SYS_sn4 DZT-norm
• all score normalizations improve performance
• DT-norm seems to perform better than T-norm and
  DZT-norm at minimum DCF point

NIST 02 SRE data (subset)
16
Conclusions

• validation of the new toolkit audioseg
• new baseline system performs well
• frame selection is crucial for good performance
• work on feature transformations (PCA, ICA ...)
• model adaptation on test data
• hierarchical structural model adaptation

• IRISA participation to NIST03 SRE
• Perspectives