Top-down Approach to Solving Speaker Diarization Errors in diaLogic System

Abstract

Speaker diarization, which separates continuous speech signals into utterances associated to different speakers, is critical to any environment that supports team collaboration using models based on data extracted from speech. However, the occurring diarization errors are hard to reduce only through better processing. This paper proposes top-down error correction based on Bayesian prediction about the most likely author of an utterance. Experiments studied the effectiveness of the method.