Scholarlink

This research introduces a modified HuBERT model for the speaker diarization process, comprising an audio signal processing framework with pre-processing, embedding extraction, and clustering stages. The system, leveraging the modified HuBERT model for speech embedding, excels at extracting high-quality representations from audio signals. Spectral clustering is employed for speaker discrimination, leading to significant improvements in efficiency compared to traditional systems. The framework is evaluated using an in-house multilingual-multispeaker dataset of eight languages and 100 audio samples, divided into training, development, and testing subsets. A comparative analysis between the HuBERT model and other methods, such as Wav2Vec2 and Gaussian Mixture Models (GMM), is conducted. Results show that HuBERT consistently outperforms these methods, achieving 15.04% diarization error rate (DER) for single-speaker scenarios and 50.16% for multi-speaker cases. In contrast, Wav2Vec2 yields 54.39% and 94.20%, while GMM produces 42.13% and 84.03% DER for single and multi-speaker scenarios, respectively.

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

Comparative Analysis of Speaker Diarization Results on Multi-lingual Multi Speaker Data and Single Speaker Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • S. Sivaprasath,
  • Aaron Samuel,
  • N. Muhammad Fadil,
  • C. Sherin Shibi

摘要

This research introduces a modified HuBERT model for the speaker diarization process, comprising an audio signal processing framework with pre-processing, embedding extraction, and clustering stages. The system, leveraging the modified HuBERT model for speech embedding, excels at extracting high-quality representations from audio signals. Spectral clustering is employed for speaker discrimination, leading to significant improvements in efficiency compared to traditional systems. The framework is evaluated using an in-house multilingual-multispeaker dataset of eight languages and 100 audio samples, divided into training, development, and testing subsets. A comparative analysis between the HuBERT model and other methods, such as Wav2Vec2 and Gaussian Mixture Models (GMM), is conducted. Results show that HuBERT consistently outperforms these methods, achieving 15.04% diarization error rate (DER) for single-speaker scenarios and 50.16% for multi-speaker cases. In contrast, Wav2Vec2 yields 54.39% and 94.20%, while GMM produces 42.13% and 84.03% DER for single and multi-speaker scenarios, respectively.