Score-level compensation of extreme speech duration variability in speaker verification

Sergio Perez-Gomez, Daniel Ramos, Javier Gonzalez-Dominguez, Joaquin Gonzalez-RodriguezInterspeech 2010

TL;DR

This paper tackles the significant performance degradation in speaker verification systems caused by varying speech utterance lengths. They propose novel score-level compensation methods, including length-conditional Gaussian modeling and logistic regression, which effectively mitigate these duration-induced misalignments, even outperforming standard normalization techniques like T-Norm.

What Problem Does This Paper Address?

Speaker verification systems aim to identify individuals based on their voice. A major hurdle in achieving robust performance is the inherent variability in speech signals, stemming from various factors like background noise, channel conditions, and crucially, the length of the speech utterance itself. While much research has focused on compensating for other session variabilities, the impact of speech duration variability, especially in scenarios like forensic applications where test segment lengths can be highly unpredictable, has been less explored. This variability leads to score misalignments, where scores generated from utterances of different lengths don't align properly, globally degrading system performance and complicating decision-making processes. Existing normalization techniques, like T-Norm, often fall short of completely resolving these duration-related issues, particularly when using fixed cohorts.

Methods

The authors propose several novel methods for compensating speech duration variability at the score level, operating within the fusion/calibration stage of speaker verification systems. These methods aim to transform scores into log-likelihood-ratios, making them interpretable under a Bayesian framework.

  • Length-conditional Gaussian Modelling (GM): This approach models the score distributions as Gaussian. It transforms scores based on the lengths of both the enrolled model speech and the test segment. The paper explores both 1D-GM (using only test segment length) and 2D-GM (using both enrollment and test segment lengths).
  • Length-conditional Linear Logistic Regression (LLR): Similar to GM, this method uses logistic regression instead of Gaussian densities to model scores. It also offers 1D-LLR and 2D-LLR variants, leveraging the lengths of the speech segments.
  • Bilinear Logistic Regression (BLR): This method extends LLR by optimizing a logistic regression model conditioned on auxiliary information, specifically utterance lengths. It introduces weights derived from normalized durations of model and test segments, exploring three variations (BLR-1, BLR-2, BLR-3) based on how these durations are incorporated.

All proposed methods are supervised and require training on a separate development set.

Datasets

  • DurTelSRE06: A newly generated corpus for training the compensation algorithms. It's derived from NIST SRE 2006 telephone-only speech data, with utterance lengths truncated to simulate variability (3, 10, 15, 20, 30, 40, 50, 60, 100, and 150 seconds).
  • DurTelSRE08: A newly generated corpus for testing the compensation algorithms. It's derived from NIST SRE 2008 telephone-only speech data, with similar length truncation as DurTelSRE06.

The paper also references standard NIST SRE 2006 and 2008 datasets for comparison.

Key Insights and Ideas

  • Duration Variability is a Major Performance Killer: Even when other session variabilities are controlled, extreme duration variability leads to a dramatic decrease in speaker verification performance, as evidenced by the significant increase in EER on the custom corpora compared to standard NIST SRE datasets.
  • T-Norm Isn't Enough: Standard score normalization techniques like T-Norm, even when applied, do not fully eliminate the performance degradation caused by duration variability, primarily due to the use of fixed cohorts. This highlights the need for more specialized compensation methods.
  • Leveraging Both Enrollment and Test Lengths is Crucial: The 2D methods (2D-GM and 2D-LLR), which utilize information from both the enrolled model's speech length and the test segment's length, consistently outperform their 1D counterparts. This suggests that the duration of the reference voice sample also plays a significant role in score misalignment.
  • Novel Methods Offer Significant Gains: The proposed GM and LLR-based compensation methods, particularly the 2D variants, demonstrate substantial improvements in EER, even when T-Norm is already applied. This indicates their effectiveness in addressing the residual duration effects.

Key Results

The paper reports significant improvements in Equal Error Rate (EER) for the DurTelSRE08 dataset when applying the proposed compensation methods.

  • Overall Improvement: For "All Durations" on DurTelSRE08, 1D-GM achieved a 4.30% EER improvement, while 2D-GM achieved 16.37%. Similarly, 1D-LLR showed 4.20% improvement, and 2D-LLR achieved 16.62%.
  • 2D Methods Superiority: The 2D methods consistently provided larger EER improvements across various test segment lengths compared to their 1D counterparts. For example, when comparing 3s and 100s test segments, 2D-GM yielded a 16.99% improvement, significantly higher than 1D-GM's 9.61%.
  • BLR Performance: While BLR methods also showed improvements, they generally underperformed compared to the GM and LLR approaches, suggesting that the larger number of parameters in GM and LLR allows for better adaptation.
  • Computational Efficiency: The LLR and GM schemes (both 1D and 2D) were noted to be superior to BLR in terms of computational cost and convergence.

Limitations and Future Work

  • Limitations: The paper acknowledges that the assumption of Gaussianity for score distributions, while supported by T-Norm usage, might not hold universally. The proposed methods are supervised and require training data. The BLR methods, while elegant, showed less performance than GM and LLR.
  • Future Work: The authors suggest a more comprehensive study of the proposed algorithms under different score normalization techniques, including adaptive schemes. They also propose exploring more advanced interpolation techniques for handling potentially continuous speech variability.

Summary updated

BibTeX
@inproceedings{perezgomez2010score,
  author = {Sergio Perez-Gomez and Daniel Ramos and Javier Gonzalez-Dominguez and Joaquin Gonzalez-Rodriguez},
  title = {{Score-level compensation of extreme speech duration variability in speaker verification}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure