Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

SSPS: Self-Supervised Positive Sampling for Robust Self-Supervised Speaker Verification

Created by
  • Haebom

Author

Theo Lepage, Reda Dehak

Outline

In this paper, we propose a novel positive sampling technique, Self-Supervised Positive Sampling (SSPS), to improve the performance of Self-Supervised Learning (SSL) in Speaker Verification (SV). Existing SSL methods have the limitation that they mainly encode recording environment information by using the same utterance of the same speaker as a positive sample. SSPS solves this problem by finding positive samples with different recording environments of the same speaker in the latent space by utilizing clustering and positive embedding memory cues. By applying SSPS to SimCLR and DINO models on the VoxCeleb1-O dataset, we achieved EER (Equal Error Rate) of 2.57% and 2.53%, respectively, which surpassed the previous best performance. In particular, SimCLR-SSPS showed similar performance to DINO-SSPS by reducing the within-speaker variance and reducing the EER by 58%.

Takeaways, Limitations

•
Takeaways:
◦
A novel method to effectively solve the __T331__ of conventional positive sampling in speaker authentication based on self-supervised learning is presented.
◦
Proposing an efficient positive sample retrieval strategy using clustering and memory queues.
◦
Demonstrates performance improvements in both SimCLR and DINO models, suggesting applicability to various SSL models.
◦
A mechanism for improving performance by reducing intra-speaker variance is presented.
◦
Achieving SOTA performance on VoxCeleb1-O dataset.
•
Limitations:
◦
The effectiveness of the proposed SSPS technique may be limited to specific datasets (VoxCeleb1-O) and models (SimCLR, DINO). Generalization performance verification for other datasets or models is required.
◦
Additional optimization studies are needed on parameter settings of clustering and memory queues.
◦
Lack of analysis of SSPS's computational cost and memory usage. Additional considerations are needed for practical system applicability.
👍
Made with Slashpage