Segmentation of telephone speech based on speech and non-speech models

Abstract

In this paper we investigate the automatic segmentation of recorded telephone conversations based on models for speech and non-speech to find sentence-like chunks for use in speech recognition systems. Presented are two different approaches, based on Gaussian Mixture Models (GMMs) and Support Vector Machines (SVMs), respectively. The proposed methods provide segmentations that allow for competitive speech recognition performance in terms of word error rate (WER) compared to manual segmentation. © 2013 Springer International Publishing.