用球面贝叶斯聚类简化端到端说话人分离，无需预训练参数。
SphereVBx: Spherical Variational Bayes Clustering for Simplified EEND-VC Diarization
- 基于环面概率球面判别分析构建球面变分贝叶斯聚类框架。
- 在多个语音分离基准上提升级联流程聚类准确率，等效或超越EEND-VC表现。
- 无需预训练参数，直接使用余弦相似度，部署更简单适合实际应用。
我们提出SphereVBx，一种基于环面概率球面判别分析（T-PSDA）的超球嵌入贝叶斯聚类框架。该方法沿用VBx的变分贝叶斯形式，但将后端的高斯概率线性判别分析（PLDA）替换为T-PSDA，实现冯·米塞斯-费舍尔分布混合模型的变分推断。我们将SphereVBx应用于说话人分离，特别是端到端神经分离与向量聚类（EEND-VC）框架。其中无参数变体SphereVBx-PF对应于球面相似性模型，与余弦打分密切相关，且无需预训练后端参数。多组语音分离基准实验表明，SphereVBx在级联式分离流水线中提升了聚类准确率，并在EEND-VC框架中实现相当或更优性能，同时显著简化了聚类阶段。
原文摘要 · Abstract (English)
We propose SphereVBx, a Bayesian clustering framework for hyperspherical embeddings based on Toroidal Probabilistic Spherical Discriminant Analysis (T-PSDA). The method follows the variational Bayesian formulation of VBx while replacing the Gaussian Probabilistic Linear Discriminant Analysis (PLDA) backend with T-PSDA, resulting in variational inference in a mixture of von Mises-Fisher distributions. We apply SphereVBx to speaker diarization and in particular to the end-to-end neural diarization with vector clustering (EEND-VC) framework. A parameter-free variant, denoted SphereVBx-PF, corresponds to a spherical similarity model closely related to cosine scoring and does not require pretrained backend parameters. Experiments on multiple diarization benchmarks show that SphereVBx improves clustering accuracy in cascaded diarization pipelines and achieves comparable or better performance in the EEND-VC framework while significantly simplifying its clustering stage.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。