Nonlinear Dynamical Invariants and SpeechSpeaker Recognition

1
Nonlinear Dynamical Invariants and Speech/Speaker
Recognition
Saurabh Prasad and Sundararajan
Srinivasan Intelligent Electronic Systems Human
and Systems Engineering Department of Electrical
and Computer Engineering
2

• Outline of Experiments

Time-Delay Embedding
Filtered-Delay Embedding
Embedding
SVD Embedding
Fractal Dimension
Lyapunov Exponent(s)
Correlation Entropy
Hurst Exponent

• Phoneme Level
• Attractors
• Variation across
• sustained phonemes
• Variation across
• speakers
• Phoneme
• recognition

• Word Level
• Attractors
• Isolated Digit
• Recognition
• TI-DIGITS
• Experiment

• Speaker Level
• Attractors
• Speaker
• Recognition tasks

3

• Subtle Issues

• Choice of Time Delay
• Using a Pitch tracking algorithm in the
  front-end?
• Weighing the Nonlinear Features Separately in
  the GMM framework
• Handling Silence Frames
• Ignore information from invariants for these
  frames?
• Mean (and Variance) Normalization