Large margin Gaussian mixture models for speaker identification

Reda Jourani, Khalid Daoudi, Régine André-Obrecht, Driss AboutajdineInterspeech 2010

TL;DR

This paper introduces a simplified Large Margin Gaussian Mixture Model (LM-GMM) for speaker identification. By focusing on optimizing only the mean vectors of the Gaussian components, their approach achieves competitive accuracy with significantly faster training times compared to existing LM-GMM methods and traditional GMMs.

What Problem Does This Paper Address?

Speaker identification systems aim to determine the identity of a speaker from a speech sample by comparing it against a set of known speaker models. Traditional methods often rely on Gaussian Mixture Models (GMMs) trained using Maximum Likelihood Estimation (MLE). While effective, MLE doesn't directly optimize for classification accuracy. Discriminative approaches, like Support Vector Machines (SVMs) combined with GMMs, have shown promise but can be computationally intensive. The paper addresses the need for a discriminative approach that is both accurate and computationally efficient for speaker identification, building upon the concept of Large Margin GMMs (LM-GMMs) which were previously applied to speech recognition but not speaker identification.

Methods

The authors propose a simplified version of the Large Margin GMM (LM-GMM) algorithm. The original LM-GMM models each class (speaker) with a mixture of ellipsoids, defined by mean vectors, positive semidefinite matrices (orientation), and scalar offsets. The proposed simplification removes the need to model the orientation matrices, focusing solely on optimizing the mean vectors. This simplification leads to a more tractable optimization problem. They also incorporate a strategy to handle outliers by re-weighting hinge loss terms based on their violation of the large margin constraints. The optimization is performed using the L-BFGS algorithm.

Datasets

The experiments were conducted using data from the NIST-SRE'2004 and NIST-SRE'2006 campaigns. A male-dependent UBM was trained on telephone data from NIST-SRE'2004. For testing, 50 male target speakers from the NIST-SRE'2006 primary task were used, with 11,600 trials.

Key Insights and Ideas

  • Simplification for Efficiency: The core insight is that optimizing only the mean vectors of the Gaussian components in an LM-GMM framework is sufficient to achieve good performance in speaker identification, while drastically reducing computational complexity.
  • Discriminative Power without Kernels: LM-GMMs, even in their simplified form, offer a discriminative approach that directly learns class boundaries in the input space, avoiding the need for computationally expensive kernel matrices common in methods like SVMs.
  • Outlier Robustness: The proposed outlier handling mechanism, by re-weighting hinge loss terms, demonstrates the importance of addressing noisy data in discriminative training for improved robustness.
  • Potential for Large-Scale Deployment: The significant reduction in training time makes this simplified LM-GMM approach highly suitable for large-scale speaker identification tasks, such as those encountered in NIST-SRE evaluations.

Key Results

The simplified LM-GMM approach achieved competitive speaker identification accuracy. Specifically, it performed as well as the original LM-GMM with 32 Gaussian components and significantly outperformed both the traditional GMM and the original LM-GMM when using 16 Gaussian components. The authors note that their simplified algorithm requires fewer iterations to converge compared to the original LM-GMM.

Limitations and Future Work

The authors acknowledge that they did not use development data for tuning the hyperparameter $\alpha$ in the LM-GMM systems, setting it to 0 for their experiments. They suggest that performance could potentially improve with the availability of development data. Future work includes performing experiments on full NIST-SRE tasks, comparing their approach to generative GMMs and GMM-supervector SVMs, and integrating Latent Factor Analysis into their training and testing procedures.

Summary updated

BibTeX
@inproceedings{jourani2010large,
  author = {Reda Jourani and Khalid Daoudi and Régine André-Obrecht and Driss Aboutajdine},
  title = {{Large margin Gaussian mixture models for speaker identification}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure