Speaker Diarization Using Linguistic Information

This technical report discusses the role of linguistic information in the process of speaker diarization, particularly in the context of automatic broadcast news transcripts. The document outlines how broadcast news programs provide valuable information that can help identify speakers and their roles, such as anchors and reporters. It details the linguistic patterns developed to detect speaker transitions and how these patterns are applied to both manual and automatic transcripts to evaluate their effectiveness. The report presents experiments comparing the performance of linguistic patterns on manual versus automatic transcriptions, highlighting the differences in accuracy and error rates. It includes specific data on the performance of self-speaker, next-speaker, and previous-speaker rules, demonstrating the impact of transcription quality on diarization outcomes. The findings indicate that while linguistic patterns can enhance speaker identification, the quality of the transcription significantly influences the overall effectiveness of the diarization process.