Contents

• Contents
• INTRODUCTION
  • Introduction to the problem.
  • Introduction to this thesis project.
  • Contents of the thesis
• BACKGROUND
  • Self-Organizing Map
    • The SOM algorithm
    • About the ordering and convergence
      • Ordering of the SOM units.
      • Asymptotic density of the SOM units.
    • Some SOM applications and properties
  • Learning Vector Quantization
    • LVQ algorithms
      • LVQ2.
      • LVQ3.
    • Minimization of classification errors
      • Measuring the misclassification.
      • Why LVQ?
  • Hidden Markov Models
    • Description of the model
      • Assumptions and definitions.
      • Probability of the sequences.
      • Output density models.
      • HMM/ANN hybrids.
    • Limitations and gains
  • Overview of speech recognition applications
    • ASR performance.
    • State-of-art in ASR error rates.
    • Examples of existing applications.
• RESULTS
  • About the phoneme recognition experiments
    • Recognition error rate.
    • ASR system.
  • The applications of SOM for MDHMMs
    • Motivation
    • Mixture density initialization by SOM
    • Segmental SOM training
      • The MDHMM training procedure
      • Convergence.
    • Characteristics of the mixture density approximation by SOM
      • Smoothing.
      • Tolerance for data defects.
      • Selective accuracy.
      • Fast computation
    • Examples of fast search methods in ordered codebooks
      • The topological K-best search
      • Some important topics for fast search.
      • Some experimental comparisons.
      • The tree-search SOM
  • LVQ for MDHMMs
    • Motivation
    • LVQ for training vector quantization codebooks
    • Corrective tuning based on LVQ2
    • Segmental LVQ3 training
    • Relations to other corrective training algorithms
• CONCLUSIONS
• References