Text%20independent%20speaker%20identification%20in%20multilingual%20environments

1
Text independent speaker identification in
multilingual environments

• I. Luengo, E. Navas, I. Sainz, I. Saratxaga,
• J. Sanchez, I. Odriozola and I. Hernaez

2
Contents

• Introduction
• SR in language mismatched conditions
• Existent solutions
• Proposed solution
• Working database
• Variability measures
• Experimental results
• Conclusions

3
Speaker Recognition System
TRAIN
Feature Extr.
Train
TEST
Accuracy decreases
Language mismatch?
Feature Extr.
Score
Decision
4
Existent solutions

• Multi-language training
• One model trained with various languages (per
  speaker)
• Model learns characteristics of different
  languages
• Multi-model training
• One model for each language (per speaker)
• Language detector

5
Existent solutionsDrawbacks

• Possible languages must be known in advance for
  each speaker
• Not generalizable for languages not seen during
  training
• More recording sessions needed for training
• Time ? Money
• Desired solution Language independent
• Suitable for languages not seen during training
• Capable of single-language training

6
Proposed solution

• Language-independent features
• Normalization?
• New features?
• Short-term intonation and energy values
• High speaker discrimination capability
• Global distribution may change little with
  language
• Combinable with MFCC
• Only in voiced frames (intonation)
• High session variability
• MVN for inter-session normalization

7
Database

• Bilingual Spanish-Basque speech database
• 22 speakers (11 Male, 11 Female)
• 4 sessions (inter-session variability)
• 7 numeric sequences (8 digits) per session and
  language

8
Variability measures

• Adding new features ALWAYS increases
  separability/variability
• Speaker separability ? discrimination
• Language variability ? model/test mismatch
• Session variability ? model/test mismatch
• Key issue Does speaker separability increase
  more than language/session variability?

9
Variability measures

• Kullback-Leibler divergence for variability
  estimation
• Interesting measures
• Good if new features increase these ratios

10
Variability measures
MFCC MFCCP Gain
Lang - 4.09 4.61 12
Spk S 6.34 8.25 30
Spk B 6.82 8.77 29
Ses S 3.62 4.81 33
Ses B 3.52 4.64 32
Spk/Lang S 1.55 1.79 15
Spk/Lang B 1.67 1.90 14
Spk/Ses S 1.75 1.72 -2
Spk/Ses B 1.94 1.89 -3
11
Experimental results
S-S B-B S-B B-S SB-S SB-B
MFCC (ref) 98.3 97.3 63.6 67.3 96.8 95.6
MFCC (V) 97.6 96.8 62.6 67.0 96.6 95.6
MFCCP (V) 97.1 96.3 71.0 73.0 96.1 94.4
Gain (V) -0.5 -0.5 13.4 9.0 -0.5 -1.3

• X-Y ? Training in X, testing in Y

12
Conclusions

• Short-term intonation and energy values increase
  language robustness
• Little accuracy drop on language-matched
  conditions
• Very useful if test language is unpredictable
• Variability measures predict results reasonably
• Allows easy selection of features prior to
  experiments

13
Text independent speaker identification in
multilingual environments

• I. Luengo, E. Navas, I. Sainz, I. Saratxaga,
• J. Sanchez, I. Odriozola and I. Hernaez