MPhil. in Computer Speech and Language Processing
Cambridge University Engineering Department

Support Vector Machines applied to Speech Pattern Classification

K. K. Chin
Darwin College

Dissertation submitted to the University of Cambridge
in partial fulfilment of the degree of Master of Philosophy

September 10, 1998

• Contents
• List of Figures
• List of Tables
• Introduction
• The Formulation of Support Vector Machine
  • Empirical Risk Minimisation
  • Structural Risk Minimisation
  • VC dimension and VC confidence
  • Linear Support Vector Machine - A Maximum Margin Classifier
  • Extending SVM to a Soft Margin Classifier
  • Nonlinear Support Vector Machine
  • SVM training and SRM
• The Training of Support Vector Machines
  • Solving the Quadratic Programming Problem
    • Comparison of training results from the three different solvers
    • Reduction of Execution Time of QP Solvers
  • Decomposition Algorithm for the SVM training
  • Incremental Training of the Support Vector Machine
  • Complexity and Scalability of the Training Algorithm
• Multi-Class Support Vector Machine
  • Combination of 1-to-rest classifiers
  • Pairwise Coupling and Pairwise with Majority Voting
  • Extending SVM to solve multi-class problems
  • Comparing the results of different approaches to multi-class problem
• Tuning the Support Vector Machine
  • Trade-off between Maximum Margin and Classification Errors
  • Selecting the Kernel function
    • Using a Radial Basis Function as kernel
    • Using a polynomial kernel
• Experiments with Speech Patterns
  • Vowel Classification - Peterson_Barney
  • Vowel Classification - TIMIT Database
  • Speaker Identification - TIMIT Database
• Conclusion and Further Work
• References
• About this document ...