Scholarlink

<p>The majority of speaker diarization systems use spectral clustering and its variants for labeling the clusters. However, the lack of incorrect initialization of the number of clusters in spectral clustering often results in poor performance of speaker diarization systems. On the other hand, a large number of speakers present in the conversation or meeting also make the performance of the diarization system critical. To deal with these issues, an improved speaker diarization pipeline is proposed. This pipeline employs a novel Robust Cluster Estimation (RCE) technique to compute an optimal number of clusters and Direct Normalized Cut (DNC) clustering to deal with the issue of a large number of speakers. The performance of the proposed pipeline method for speaker diarization is evaluated using widespread evaluation metric Diarization Error Rate (DER) as 11.85%, 13.8%, and 30.6% on benchmarked diarization datasets; VoxConverse (Test), AMI, and DISPLACE respectively. An intensive experimental evaluation is also done to compare the proposed system with state-of-the-art approaches as well as existing baselines. Results obtained have shown that the proposed method has outperformed state-of-the-art diarization systems and existing baselines. Further, a thorough analysis is done to show the effectiveness of the proposed pipeline.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Direct normalized cut clustering using a novel robust cluster estimation technique for multi-speaker diarization

  • 期刊论文

  • 发表日期：

    2025年6月1日

查看全文
  • Aishwarya Gupta,
  • Archana Purwar

摘要

The majority of speaker diarization systems use spectral clustering and its variants for labeling the clusters. However, the lack of incorrect initialization of the number of clusters in spectral clustering often results in poor performance of speaker diarization systems. On the other hand, a large number of speakers present in the conversation or meeting also make the performance of the diarization system critical. To deal with these issues, an improved speaker diarization pipeline is proposed. This pipeline employs a novel Robust Cluster Estimation (RCE) technique to compute an optimal number of clusters and Direct Normalized Cut (DNC) clustering to deal with the issue of a large number of speakers. The performance of the proposed pipeline method for speaker diarization is evaluated using widespread evaluation metric Diarization Error Rate (DER) as 11.85%, 13.8%, and 30.6% on benchmarked diarization datasets; VoxConverse (Test), AMI, and DISPLACE respectively. An intensive experimental evaluation is also done to compare the proposed system with state-of-the-art approaches as well as existing baselines. Results obtained have shown that the proposed method has outperformed state-of-the-art diarization systems and existing baselines. Further, a thorough analysis is done to show the effectiveness of the proposed pipeline.