A generalised derivative kernel for speaker verification

C. Longworth, M. J. F. GalesInterspeech 2008

TL;DR

This paper introduces a novel "generalised derivative kernel" (GDK) for speaker verification that unifies existing dynamic kernel approaches. By combining static kernels with generative models, the GDK offers a more principled way to extract features from variable-length speech utterances, leading to improved performance.

What Problem Does This Paper Address?

Speaker verification systems aim to determine if a speech utterance belongs to a claimed speaker. Support Vector Machines (SVMs) are powerful classifiers for this task, but they typically require fixed-dimensional input. Speech utterances, however, are variable-length sequences of observations. This mismatch necessitates the use of "dynamic kernels" that can map these sequences into a fixed-dimensional space. Existing dynamic kernels, like the GLDS kernel, often average features, potentially losing information. Derivative kernels, on the other hand, use generative models but operate on original observation spaces. This paper bridges this gap by proposing a dynamic kernel that leverages a static kernel's feature space for derivative calculations, offering a more robust and principled feature extraction method.

Methods

The core contribution is the Generalised Derivative Kernel (GDK). The GDK extends the concept of derivative kernels by performing derivative calculations not in the original observation space, but in the feature space defined by a static kernel. This allows for richer feature representations.

Key aspects of the method include:

  • Combining Static and Dynamic Kernels: A static kernel (e.g., polynomial, Gaussian) is first applied to individual observations to map them into a higher-dimensional feature space. Then, a generative model (like a Gaussian Mixture Model - GMM) is defined within this feature space. Derivatives of the log-likelihood with respect to the parameters of this feature-space GMM form the new features.
  • Generalization: The GDK is shown to generalize both the GLDS kernel (when the static kernel is linear and the generative model is a single component GMM) and standard derivative kernels.
  • Approximations for Practicality: Training a generative model directly in the high-dimensional static kernel feature space is often infeasible. The paper proposes approximations:
  • Component Posterior Approximation: Instead of computing posteriors in the feature space, posteriors from a model trained in the original observation space are used.
  • Static Kernel Estimation Approximation: The variance and metric terms of the feature-space GMM are approximated, often by assuming an identity metric and using a form derived from the original static kernel. This leads to a kernel that can be computed without explicitly generating the feature-space representations.

Datasets

The proposed GDK was evaluated on the NIST 2002 SRE (Speaker Recognition Evaluation) dataset. This dataset consists of cellular speech data from 139 male and 191 female speakers. The paper notes that their experimental setup did not strictly adhere to the NIST SRE protocol due to limited development data, using enrollment data for both UBM training and imposter modeling.

Key Insights and Ideas

  • Bridging the Gap: The GDK elegantly unifies two previously distinct approaches to dynamic kernels (GLDS and derivative kernels), providing a theoretical framework that explains their relationship.
  • Feature Space Matters: Performing derivative calculations in a richer, static-kernel-induced feature space can capture more discriminative information than operating solely on raw observations.
  • Approximation is Key: The paper highlights the critical role of approximations in making these powerful kernel ideas computationally tractable for real-world applications. The choice of approximation can significantly impact performance.
  • Non-linear Static Kernels Improve Performance: Using non-linear static kernels (like polynomial) within the GDK framework led to performance gains, suggesting that the richer feature spaces they create are beneficial.

Key Results

  • SVM Outperforms LLR: In general, SVM classifiers using the GDK outperformed traditional GMM-LLR classifiers.
  • GDK Performance: The GDK achieved competitive results, with the best performance obtained using a second-order polynomial static kernel with a 16-component model, yielding an Equal Error Rate (EER) of 9.25%.
  • Impact of Approximations:
    • Using Viterbi alignments (a form of hard component alignment) generally degraded performance compared to explicit models.
  • Approximating component posteriors using linear models (from the observation space) improved performance for polynomial kernels, suggesting robustness.
  • Model Size Trade-off: For linear kernels, performance improved with larger model sizes (up to 1024 components). However, for polynomial kernels, smaller models (16 components) performed best, likely due to the increased dimensionality of the feature space and potential over-training with larger models on the given dataset.

Limitations and Future Work

  • Approximation Robustness: The paper acknowledges that approximations for component posteriors and static kernel estimation might not be robust when the relative distances between observations differ significantly between the observation and feature spaces.
  • Dataset Size: The authors suggest that with more training data, further gains might be achieved using larger model sizes, particularly for polynomial kernels.
  • Computational Cost: While approximations are used, the computation of the GDK, especially with non-linear static kernels and large datasets, can still be computationally intensive. Future work could focus on further optimizing these computations.
  • Exploration of Static Kernels: The paper explored linear, polynomial, and Gaussian static kernels. Further investigation into other types of static kernels or combinations of kernels (e.g., using Multiple Kernel Learning) could yield additional improvements.

Summary updated

BibTeX
@inproceedings{longworth2008generalised,
  author = {C. Longworth and M. J. F. Gales},
  title = {{A generalised derivative kernel for speaker verification}},
  booktitle = {Interspeech},
  year = {2008},
}
Figure