Scholarlink

This paper presents the AuroraLab system for the Voice Timbre Attribute Detection (vTAD) 2025 Challenge. In this challenge, we propose a novel framework that introduces discriminative speaker embeddings (DSE) into the vTAD task, termed DSE-vTAD. DSE-vTAD leverages strong speaker embedding extractors to obtain discriminative speaker embeddings. In addition, unlike the challenge’s baseline system, DSE-vTAD concatenates a pair of speaker embeddings along with their Hadamard product features. Compared with the baseline system, DSE-vTAD achieves significant performance improvements. On the unseen test set, the best DSE-vTAD system achieves 91.31% Avg ACC and 8.54% Avg EER. On the seen-speaker test set, the best DSE-vTAD system achieves 97.32% Avg ACC and 2.72% Avg EER. The source code is released at https://github.com/zds-potato/DSE-vTAD .

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

Introducing Discriminative Speaker Embeddings for Voice Timbre Attribute Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhida Song,
  • Liang He

摘要

This paper presents the AuroraLab system for the Voice Timbre Attribute Detection (vTAD) 2025 Challenge. In this challenge, we propose a novel framework that introduces discriminative speaker embeddings (DSE) into the vTAD task, termed DSE-vTAD. DSE-vTAD leverages strong speaker embedding extractors to obtain discriminative speaker embeddings. In addition, unlike the challenge’s baseline system, DSE-vTAD concatenates a pair of speaker embeddings along with their Hadamard product features. Compared with the baseline system, DSE-vTAD achieves significant performance improvements. On the unseen test set, the best DSE-vTAD system achieves 91.31% Avg ACC and 8.54% Avg EER. On the seen-speaker test set, the best DSE-vTAD system achieves 97.32% Avg ACC and 2.72% Avg EER. The source code is released at https://github.com/zds-potato/DSE-vTAD .