Margin Based Learning Framework with Geometric Margin Minimum Classification Error for Robust Speech Recognition

Abstract

Statistical learning theorycombines empirical risk and generalization functionin single optimized objective function of margin based learning for optimization. Margin concept incorporating in Hidden Markov Model (HMM)for speech recognition, Margin based learning frame work based on minimum classification error (MCE) training criteria show higher capability over any other conventional DT methods in improvingclassification robustness (generalization capability) of the acoustic model by increasing the functional margin of the acoustic model. This paper introduces Geometric Margin based separation measure in the loss function definition of margin based learning frame work instead of functional margin separation measure to develop a mathematical framework of new optimize objective function of soft margin estimation (SME) for ASR. Derived SME objective function based on Geometric Margin based separation (misclassification) measure would be capable for representing the strength of margin based learning framework in term of classification robustness by minimizing the classification error probability as well asmaximizing the geometric margin.