A GMM-based probabilistic sequence kernel for speaker verification

Kong-Aik Lee, Changhuai You, Haizhou Li, Tomi Kinnunen

Introduction

Speaker verification, the biometric task of authenticating a person's identity via their voice, has traditionally relied on two distinct paradigms: generative modeling and discriminative classification. Generative approaches, such as the Gaussian Mixture Model-Universal Background Model (GMM-UBM), excel at modeling the underlying distribution of speech features but do not explicitly optimize for the decision boundary between speakers. Conversely, discriminative methods like Support Vector Machines (SVMs) are designed to maximize class separation but typically require fixed-length input vectors, posing a challenge for variable-length speech utterances.

This paper introduces a hybrid approach that bridges this gap using a probabilistic sequence kernel. By leveraging GMM components as nonlinear basis functions, the system transforms variable-length speech into a fixed-length "characteristic vector" suitable for SVM classification.

Methodology: The Probabilistic Sequence Kernel

The proposed system utilizes a GMM as a feature extractor. The core innovation lies in the construction of a feature space that is inherently "speech-aware."

Workflow

The process of generating a characteristic vector from a speech utterance $X$ involves several stages:

  1. Basis Function Definition: A pool of $2M$ Gaussian densities is established. This pool consists of $M$ components from a Universal Background Model (representing general human speech) and $M$ components adapted specifically to the target speaker.
  2. Nonlinear Mapping: Each frame of the utterance is mapped through these $2M$ Gaussians to produce a probabilistic alignment vector, where each element represents the probability that a specific phonetic class was present.
  3. Averaging and Normalization: To handle variable utterance lengths, these alignment vectors are averaged over the entire sequence. The resulting vector is then normalized using a correlation matrix to ensure statistical stability.
The workflow for extracting characteristic vectors using GMM-based basis functions.

This architecture allows the SVM to operate in a high-dimensional space where the dimensions correspond to broad phonetic classes (e.g., vowels or nasals). This "phonetic awareness" enables the classifier to weight sounds that are most discriminative for a particular speaker.

Experimental Evaluation

The system was evaluated using the 2001 NIST Speaker Recognition Evaluation (SRE) corpus, which features cellular telephone recordings characterized by significant channel noise and variability. The evaluation involved 174 speakers and over 22,000 trials. Speech features were represented as 36-dimensional vectors comprising 12 Mel-Frequency Cepstral Coefficients (MFCCs) along with their first and second derivatives.

Results and Discussion

The performance of the proposed probabilistic sequence kernel was compared against two benchmarks: a standard GMM-UBM system and an SVM using a Generalized Linear Discriminant Sequence (GLDS) kernel (polynomial expansion).

The results, summarized by the Equal Error Rate (EER), demonstrate the superiority of the hybrid approach:

  • GMM-UBM: 8.78% EER
  • GLDS SVM: 8.49% EER
  • Probabilistic Kernel SVM: 8.10% EER
DET curves comparing the performance of GMM-UBM, GLDS SVM, and the proposed Probabilistic Sequence Kernel SVM.

As shown in the Detection Error Tradeoff (DET) curve, the proposed method is particularly effective at low False Acceptance Rates, making it highly suitable for high-security applications. Furthermore, the probabilistic kernel achieves these results with a significantly lower dimensionality (1024 dimensions) compared to the GLDS kernel (over 9000 dimensions), highlighting its efficiency in feature representation.

Conclusion

The GMM-based probabilistic sequence kernel successfully combines the strengths of generative and discriminative modeling. By using Gaussian components as speech-informed basis functions, the system provides a more meaningful and compact feature space for SVMs. While the current implementation assumes a diagonal correlation matrix for efficiency, future work may explore full-covariance normalization to further enhance performance.

Summary updated

BibTeX
@inproceedings{lee2007gmm,
  author = {Kong-Aik Lee and Changhuai You and Haizhou Li and Tomi Kinnunen},
  title = {{A GMM-based probabilistic sequence kernel for speaker verification}},
  booktitle = {Interspeech},
  year = {2007},
  doi = {10.21437/interspeech.2007-131},
}
Figure