Spoken Language Understanding in Embedded Systems

Authors: Karl Weilhammer, Prince Kumar, Volker Springer, Dominique Massonie

Abstract:

We investigated classification with Support Vector Machines for spoken language understanding with respect to their use in embedded devices, which are often equipped with slow CPUs, and main or persistent memory of limited size or with slow access times. We started with uni- and bigrams as features and managed to reduce the feature set in most cases by applying Recursive Feature Elimination from a few thousands to a few dozens. This corresponds to a reduction of the overall model size to 6% of the original size, without having a substantial loss in classification performance. The F score difference is 0.16%.