A Copula-Based Generative Score-Level Fusion Model for Speaker Verification

Sandro CumaniInterspeech 2025

TL;DR

This paper introduces a novel generative approach for combining scores from multiple speaker verification systems. By using copulas to model the joint distribution of scores, the method effectively separates the dependency structure between systems from their individual score characteristics, leading to more accurate and well-calibrated fusion results.

What Problem Does This Paper Address?

Speaker verification systems, which aim to identify individuals based on their voice, often benefit from combining the outputs of multiple individual systems. This "score-level fusion" is a flexible way to improve overall accuracy. Traditional methods often rely on linear combinations of scores, which can be suboptimal. While generative models have shown promise for calibrating individual system scores, extending them to effectively fuse multiple systems has been challenging. The core problem is how to model the complex, multi-dimensional distribution of scores from several systems in a way that is both accurate and computationally tractable, especially when individual systems might have complex score distributions.

Methods

The paper proposes a copula-based generative score-level fusion model. The key idea is to leverage Sklar's theorem, which allows for the decomposition of a joint probability distribution into its marginal distributions and a copula function that models the dependencies between them.

Here's a breakdown:

  • Marginal Models: The authors use the Variance-Gamma (VF) distribution to model the target (same speaker) and non-target (different speaker) score distributions for each individual system. The VF distribution is chosen for its flexibility in capturing complex score characteristics, outperforming simpler Gaussian models.
  • Dependency Model (Copula): A Gaussian copula is used to model the joint dependency structure between the scores of different systems. This copula captures how the scores of various systems co-vary.
  • Decoupling: The copula approach elegantly decouples the modeling of individual system score distributions (marginals) from the modeling of how these scores relate to each other (dependency). This allows for the use of sophisticated marginal models (like VF) without making the overall joint model intractable.
  • Inference Function for Margins (IFM): To estimate the model parameters, a two-step approach is used. First, the parameters for the individual VF marginal models are estimated. Second, the parameters for the Gaussian copula are estimated.

Datasets

  • Training/Development: NIST SRE 2019 Progress set was used for estimating marginal and copula parameters.
  • Evaluation:
    • NIST SRE 2019 Evaluation set.
  • Speakers in the Wild (SITW) Evaluation set.

Key Insights and Ideas

  • Copulas as a Unifying Framework: The paper demonstrates that copulas provide a powerful and flexible way to combine generative calibration models for score-level fusion. They allow for the separation of marginal and dependency modeling, which is crucial for handling complex score distributions.
  • Variance-Gamma for Richer Marginals: Using the Variance-Gamma distribution for marginal score modeling significantly improves accuracy compared to simpler models like Gaussian, as it better captures the nuances of speaker verification scores.
  • Generative Fusion Outperforms Discriminative Baselines: The proposed generative approach, by accurately modeling the underlying score distributions, achieves competitive or superior performance compared to state-of-the-art discriminative fusion methods like Logistic Regression.
  • Well-Calibrated Scores: A key benefit of generative models is their ability to produce well-calibrated scores, meaning the confidence associated with a score accurately reflects the true probability of a correct identification. This is maintained and improved in the fusion model.

Key Results

  • The proposed VF-GC (Variance-Gamma Gaussian Copula) model achieved competitive results on both NIST SRE 2019 and SITW datasets.
  • On NIST SRE 2019, the VF-GC model showed a reduction in the calibration-sensitive Cur metric of up to 7% relative to discriminative linear fusion methods.
  • The model consistently achieved optimal or near-optimal Cprim (NIST's primary cost metric) and EER (Equal Error Rate), indicating strong performance across different evaluation criteria.
  • The VF-GC model outperformed a Gaussian copula fusion model (CMLG-GC), highlighting the importance of the flexible VF marginals.

Limitations and Future Work

  • Computational Cost of VF CDF: The paper notes that the Variance-Gamma CDF does not have a closed-form expression, requiring numerical integration, which adds to the computational cost.
  • Future Work: The authors suggest exploring more complex copula functions beyond the Gaussian copula. This could potentially capture more intricate dependency structures between system scores and further improve fusion performance. They also mention the potential for analyzing more complex marginal distributions.

Summary updated

BibTeX
@inproceedings{cumani2025copula,
  author = {Sandro Cumani},
  title = {{A Copula-Based Generative Score-Level Fusion Model for Speaker Verification}},
  booktitle = {Interspeech},
  year = {2025},
}
Figure