Performance Evaluation of HOARD Framework for Podcast Application for Multi-Speaker Diarization

Nowadays Podcasts are in use by various domains to grow the audience engagements around the world for diverse application. As podcast conversation are long and contain overlapping speakers. It necessitates the need for effective speaker diarization pipeline that provides solution for “who spoke when?” problem. Few researchers have suggested various solutions for podcast speaker diarization pipeline. These work do not handle issues like large number of speakers, overlapping speakers, background and environmental noise. To overcome these challenges recently a HOARD framework has been introduced. Hence, this paper validates the effectiveness of Handling Overlap-Aware Refined Diarization (HOARD) Framework for podcast application using The American Life (TAL) podcast dataset. The framework outperformed recently proposed Overlap-Aware Multiclass Spectral Clustering in Cluster Purity and Cluster Coverage metrics. Also, reduction in Diarization Error Rate (DER) by 0.045% is observed.