Nurturing filled pause detection for spontaneous speech retrieval

Abstract

In this paper we investigate methods to adapt a system for filled pause (FP) disfluency removal to different data properties. A gradient descent algorithm for parameter optimization is presented which achieves 80.6% recall and 87.7% precision on the FP dataset and 46.5% recall and 79.6% precision on the FPElo dataset. This compares to the results produced with handoptimization on the test set. Furthermore we investigated the impact of crossvalidation and training set selection on recognizer output in order to improve the speech retrieval system.