mlsanityData to 14 Oct 2026

Understanding Hyperspherical Geometry of ECAPA-TDNN Embedding and Its Impact on Zero-Shot Voice Conversion

arXiv 2609.24688 ↗First posted 21 Sep 2026cs.SD cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

Angular-margin speaker encoders are widely used in voice conversion, yet the geometry of their classifier prototypes remains poorly understood. We analyze ECAPA-TDNN classifier prototypes as points on the unit hypersphere and characterize their organization using rotation-invariant angular statistics together with global and local effective dimensionality measures. Our analysis shows that standard training can induce angular concentration and a substantial reduction in effective dimensionality. To address this, we investigate two geometric regularization strategies (hinged Riesz log-energy and effective-dimension maximization) applied to classifier prototypes to encourage more uniform hyperspherical coverage. The resulting prototype sets exhibit higher effective dimensionality and improved isotropy, with configuration-dependent effects on speaker-recognition performance. When the corresponding ECAPA-TDNN models are used as speaker encoders for Fast-VGAN, the regularized systems also exhibit improved robustness in zero-shot voice conversion, particularly for previously unseen speakers.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 20 of its 31 references that are in the map. First posted on arXiv 21 Sep 2026.

Builds on

20 of 31 references.

References · 20

  1. 2020Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hyperspherecited by 1,405
  2. 2018CosFace: Large Margin Cosine Loss for Deep Face Recognitioncited by 971
  3. 2017SphereFace: Deep Hypersphere Embedding for Face Recognitioncited by 737
  4. 2020Prevalence of Neural Collapse during the terminal phase of deep learning trainingcited by 719
  5. 2019LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speechcited by 649
  6. 2019Intrinsic dimension of data representations in deep neural networkscited by 282
  7. 2018Implicit Self-Regularization in Deep Neural Networks: Evidence from Random Matrix Theory and Implications for Learningcited by 215
  8. 2023LibriTTS-R: A Restored Multi-Speaker Text-to-Speech Corpuscited by 101
  9. 2021Large-scale Self-Supervised Speech Representation Learning for Automatic Speaker Verificationcited by 78
  10. 2018Learning towards Minimum Hyperspherical Energycited by 75
  11. 2023Voice Conversion With Just Nearest Neighborscited by 68
  12. 2021ECAPA-TDNN Embeddings for Speaker Diarizationcited by 47
8 more
  1. 2024Zero-shot Voice Conversion with Diffusion Transformerscited by 43
  2. 2019Dimensionality compression and expansion in Deep Neural Networkscited by 38
  3. 2020The IDLAB VoxCeleb Speaker Recognition Challenge 2020 System Descriptioncited by 17
  4. 2025GenVC: Self-Supervised Zero-Shot Voice Conversioncited by 8
  5. 2018Angular Softmax Loss for End-to-end Speaker Verificationcited by 7
  6. 2022ECAPA-TDNN for Multi-speaker Text-to-speech Synthesiscited by 5
  7. 2025EZ-VC: Easy Zero-shot Any-to-Any Voice Conversioncited by 2
  8. 2025Fast-VGAN: Lightweight Voice Conversion with Explicit Control of F0 and Duration Parameterscited by 2

Built on by

Not cited yet.