Novel Method to Combine Phone-level Confidence Scores Using Support Vector Machines

Support vector machines (SVM) represents a new approach to pattern classification developed from the theory of structural risk minimization (V. Vapnik, 1998). In this paper, we propose an investigation into the application of SVM to the confidence measurement problem in speech recognition. Confidence measures are computed using the phone-level information provided by a hidden Markov model (HMM) based speech recognizer. We use support vector machines to combine phone-level confidence measures rather than traditional average techniques such as arithmetic, geometric and harmonic averages. Then a confidence measure for each word is computed by SVM and the decision of rejection or acceptance is made based on the confidence scores. Experiments of Mandarin command recognition showed that better performance can be obtained when using the proposed method