HSVM - A SVM Toolkit for Segmented Speech Data

Authors: André Stuhlsatz

Abstract:

This paper introduces add-on tools aligned with the Hidden-Markov- Model Toolkit (HTK) to use Support Vector Machines (SVM) in speech recognition. The resulting method and tool is named HSVM: Hidden-Markov-Model Toolkit using Support Vector Machines. Because SVMs have proven their generalization performance compared to the common maximum likelihood or MAP approach, a speech recognizer can profit from the use of SVMs for classifying the acoustic features. Given segmented speech data provided by a underlying HMM, the presented tools enable to post-classify acoustic features based on N-best-lists or recognition lattices. HSVM transforms the SVM predictions to probabilities and feeds them back into the recognition process in a offline fashion. Improvements using HSVM are presented for a phoneme classification on the TIMIT corpus as well as Wallstreet Journal Cambridge corpus.