Abstract

In this paper we present a set of techniques we employed in our Janus Recognition Toolkit (JRTk) Switchboard and CallHome recognizer in order to deal with imperfections in the transcriptions: inconsistent transcription of pronunciations and contractions as well as errors in utterance segmentations. These techniques consist of a dynamic, speaking mode dependent pronunciation model and a flexible utterance alignment procedure which is based on speaker adapted models (label boosting). The idea is (a) to automatically retranscribe the training corpus based on these models and procedures, (b) to train a recognizer based on these flexible transcription graphs and (c) to decode with a dynamic speaking mode dependent dictionary. The framework is successfully applied to increase the performance of our state-of-the-art JRTk Switchboard recognizer significantly. 1 Introduction Recognition of conversational speech is one of the most challenging speech recognition tasks to-date. While recognition...

Citations