Extracting true speaker identities from transcriptions

Abstract

Automatic speaker diarization generally produces a generic label such a spkr1 rather than the true identity of the speaker. Recently, two approaches based on lexical rules were proposed to extract the true identity of the speaker from the transcriptions of the audio recording without any a priori acoustic information: one uses n-gram, the other one uses semantic classification trees (SCT). The latter was proposed by the authors of this paper. In this paper, the two methods are compared in experiments carried out on French broadcast news records from the ESTER 2005 evaluation campaign. Experiments are processed on manual and automatic transcriptions. On manual transcriptions, the n-gram-based approach can be more precise, but the automatic transcriptions, the SCT-based approach gives significantly the best results in terms of recall and precision.