Effect of Relevance Factor of Maximum a posteriori Adaptation for GMM-SVM in Speaker and Language Recognition

Changhuai You, Haizhou Li, Bin Ma, Kong-Aik Lee

Introduction: The Role of MAP Adaptation

In Gaussian Mixture Model - Support Vector Machine (GMM-SVM) frameworks, Maximum a Posteriori (MAP) adaptation is the standard technique for deriving speaker-specific models from a Universal Background Model (UBM). A central parameter in this process is the relevance factor ($\gamma$), which controls the trade-off between the prior knowledge (UBM) and the observed adaptation data. Traditionally, $\gamma$ is set as a fixed constant. However, this paper demonstrates that a one-size-fits-all approach is suboptimal across diverse recording conditions and utterance durations.

The Problem with Fixed Factors

The sensitivity of system performance to the choice of $\gamma$ is evident when examining different acoustic environments. As shown in the figure below, the optimal value for $\log_2 \gamma$ varies significantly between telephone-to-telephone (tel-tel) and interview-to-telephone (int-tel) conditions. A value that minimizes the Equal Error Rate (EER) for one task may be poorly suited for another, highlighting the need for a more flexible adaptation mechanism.

EER performance across different NIST SRE 2008 conditions as a function of the relevance factor.

Beyond Fixed Factors: Data-Dependent and Adaptive Approaches

To address these limitations, the authors propose two advanced strategies:

  1. Data-dependent Relevance Factor (dep-RF): This method replaces the scalar $\gamma$ with a diagonal transfer matrix learned via an Expectation-Maximization (EM) algorithm. This allows the system to weigh Gaussian components differently based on their reliability in the training data distribution.
  2. Adaptive Relevance Factor (adp-RF): Building on dep-RF, this approach incorporates the duration of the adaptation utterance. By scaling the relevance factor according to the amount of available data, the system prevents over-fitting in short clips while maximizing information gain in longer ones.

The system further utilizes Nuisance Attribute Projection (NAP) to mitigate channel noise and employs Bhattacharyya-based kernels to facilitate effective SVM classification.

Performance Evaluation

The proposed methods were evaluated on NIST SRE 2008 and LRE 2009/2011 benchmarks. The adaptive approach (adp-RF) consistently outperformed the fixed baseline, particularly in challenging short-duration scenarios. The following figure illustrates this improvement for the 10-second evaluation task, where adp-RF achieves lower EER and minimum Detection Cost Function (DCF) values compared to both fixed and data-dependent alternatives.

Comparison of fixed, data-dependent, and adaptive relevance factors on short-duration speaker recognition tasks.

In language recognition (LRE 2009), the adaptive factor reduced the EER from 5.41% to 4.47% for 30-second utterances. These results confirm that accounting for data uncertainty and duration leads to more robust speaker and language representations.

Conclusion

By evolving the relevance factor from a static "knob" to a dynamic, duration-aware parameter, this research significantly enhanced the reliability of GMM-SVM systems. While modern deep learning architectures have largely superseded GMMs, the principle of duration-adaptive modeling remains a cornerstone of robust biometric systems.

Summary updated

BibTeX
@inproceedings{you2012effect,
  author = {Changhuai You and Haizhou Li and Bin Ma and Kong-Aik Lee},
  title = {{Effect of Relevance Factor of Maximum a posteriori Adaptation for GMM-SVM in Speaker and Language Recognition}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-551},
}
Figure