Discrimination Measure from speech signal

1
Discrimination Measure from speech signal
Seungchan Lee Department of Electrical and
Computer Engineering Mississippi State University
2

• Overview

• Speaker Recognition System
• Continue to run experiments.
• Replicate interspeech paper (2006) experiment.
• Confirm speaker-specific information from
  nonlinear invariants
• Data Transfer to ECE
• Next Plans

3

• Problem

• MFCC Nonlinear Invariants
• All experiments so far did not give better result
  than MFCC only.
• It requires to confirm speaker specific
  information from invariants.
• ? How ?
• Measures of discriminality between features
• Correlation between MFCCs and invariants
• - The values are ranged from -0.5 to 0.4 at
  MFCC features, but the
• values between MFCCs and Lyapunov
  exponents are less than the
• values of MFCC features. (less than 0.1)
• Kullback-Leibler (KL) divergence measure
• - Run the same experiments of Interspeech
  paper (2006)
• - Apply this procedure for the
  discriminality among speakers

4

• Results

• Lyapunov exponents
• Discrimination measures are high between vowels
  and fricatives and nasals and fricatives.
• Discrimination measures are small between nasals
  and vowel sounds.

5

• Results

KL-Divergence of Correlation Entropy
KL-Divergence of Correlation Dimension
6

• Results

• Lyapunov exponents
• KL-divergence of vowels of same speaker is small.
• KL-divergence of vowels of different speakers is
  high.
• Since this result is only from small dataset,
  there is no guarantee it will be good for large
  dataset.
• More investigate several possible combinations of
  dataset.
• How can overcome the overlap of different
  phonemes and different speaker?

Speaker2
Vowels
Speaker3
Speaker1
Nasals
Speaker4
fricatives
Speaker5
. . .
7

• Future Direction

• The LE only features
• Give poor result,
• But, it has certain amount of speaker-specific
  information.
• Alternatives?

Speech
MFCCs
Invariants
Scores
Scores
Combine
Or, ..
Decision
8

• Next Plan

• Nonlinear invariants for speaker recognition
• Apply the discriminality between features to the
  modeling methods.
• Be familiar with Correlation Dimension and
  Correlation Entropy class.
• Finish LE for adding features.
• Use two invariants for speaker recognition.