Variational dynamic kernels for speaker verification

C. Longworth, R. C. van Dalen, M. J. F. GalesInterspeech 2009

TL;DR

This paper introduces new "variational dynamic kernels" for speaker verification that overcome limitations of existing methods. By using more accurate approximations of the divergence between speaker models (GMMs), these kernels allow for more flexible model structures and potentially better performance, as demonstrated on the NIST 2002 SRE dataset.

What Problem Does This Paper Address?

Speaker verification (SV) systems aim to determine if a speech utterance belongs to a claimed speaker. Support Vector Machines (SVMs) are a popular choice for SV, often using "dynamic kernels" that measure similarity between generative models of speech, like Gaussian Mixture Models (GMMs). A common approach uses the Kullback-Leibler (KL) divergence between GMMs. However, calculating the exact KL divergence between GMMs is computationally intractable. Existing methods rely on approximations, such as the "matched-pair bound," which impose strict constraints: all GMMs must have the same number of components and be adapted from a single background model. This limits the flexibility of SV systems, preventing them from optimally adapting to diverse conditions like different genders, noise levels, or dialects. This paper addresses this by proposing new dynamic kernels based on more advanced variational approximations to the KL divergence, which relax these structural constraints.

Methods

The core contribution of this paper lies in deriving new dynamic kernels from variational approximations to the KL divergence between GMMs.

  • KL Divergence: The paper starts by defining the KL divergence and its form for Gaussian distributions. For GMMs, it highlights the lack of a closed-form solution and the reliance on approximations.
  • Matched-Pair Bound: This is the standard approximation used in GMM-supervector kernels. It's a simplified upper bound but requires GMMs to have the same number of components and coordinated components.
  • Variational Approximation (VAR): This method, derived by Hershey and Olsen, decomposes the KL divergence into two expected values and finds lower bounds for each using variational distributions. The resulting approximation (Equation 7) is not a strict bound but is shown to be close in practice.
  • Variational Upper Bound (UP): This approach, derived independently, also uses variational distributions to obtain a lower bound on the KL divergence (Equation 8). It involves iterative updates for the variational parameters.
  • Dynamic Kernels: The paper then proposes new dynamic kernels by exponentiating the sum of the forward and backward variational approximations of the KL divergence (Equations 15 and 16). This approach is analogous to how existing GMM-supervector kernels are formed from distance metrics. The key advantage is that these variational kernels do not require GMMs to have the same structure or be adapted from the same background model.

Datasets

  • NIST 2002 SRE dataset: This dataset was used for evaluating the performance of the proposed variational dynamic kernels. The paper notes that this dataset was chosen because it's one of the more recent publicly available evaluation datasets.

Key Insights and Ideas

  • Flexibility is Key: The most significant insight is that relaxing the structural constraints imposed by the matched-pair bound (e.g., requiring identical GMM structures) can lead to more robust and potentially better-performing speaker verification systems.
  • Variational Approximations Offer a Better Trade-off: The variational approximations (VAR and UP) provide a more accurate estimate of the KL divergence than the matched-pair bound, especially when GMM structures differ or are adapted from different background models.
  • Variational Kernels Mimic Existing Forms: By exponentiating the sum of forward and backward variational KL divergences, the authors create kernels that share a similar functional form with established dynamic kernels, making them directly applicable within SVM frameworks.
  • Cross-Gender Performance Boost: The variational kernels showed similar performance for both within-gender and cross-gender comparisons, unlike the matched-pair bound which degraded significantly in cross-gender scenarios. This suggests they are better suited for handling variations in speaker characteristics.

Key Results

  • Improved Accuracy with Variational Kernels: The proposed variational kernels (VAR and UP) achieved comparable or better performance than the standard GMM-supervector (GMM-SV) kernel on the NIST 2002 SRE task. The best performing kernel, VAR, achieved an Equal Error Rate (EER) of 9.68%.
  • Robustness to Model Structure: For cross-gender comparisons, the matched-pair bound kernel's performance degraded significantly (by an order of magnitude) compared to the true KL divergence. In contrast, the variational approximations (VAR and UP) showed similar accuracy for both cross- and within-gender comparisons.
  • Adaptability to Background Models: When using gender-independent background models, the variational kernels showed smaller performance degradation compared to the matched-pair kernel, highlighting their advantage in scenarios with less coordinated GMMs.

Limitations and Future Work

  • Normalization Not Explored: The paper mentions that normalizing the KL divergence, as done in some existing GMM-supervector kernels, was not explored for the variational approximations but could lead to further gains.
  • Polarization Identity Not Used: The authors note that deriving kernels using the polarization identity (which relates distance metrics to kernels) is difficult for the variational approximations and will be a subject of future work.
  • More Complex Training Schemes: Future work could involve exploring more complex training schemes, such as allowing GMM structures to vary or adapting clusters of speakers from a wider range of background models.
  • Variational Kernels with Polarization: Developing variational kernels derived using the polarization identity is suggested as a path for further improvement.

Summary updated

BibTeX
@inproceedings{longworth2009variational,
  author = {C. Longworth and R. C. van Dalen and M. J. F. Gales},
  title = {{Variational dynamic kernels for speaker verification}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure