Speaker recognition using supervised probabilistic principal component analysis

Yun Lei, John H. L. HansenInterspeech 2010

TL;DR

This paper introduces a supervised variant of Probabilistic Principal Component Analysis (SPPCA) for speaker recognition, aiming to better leverage speaker identity information during feature extraction. By integrating speaker labels into the model, SPPCA demonstrates improved performance over its unsupervised counterpart (PPCA) and offers a promising direction for enhancing speaker recognition systems.

What Problem Does This Paper Address?

Traditional speaker recognition systems often struggle with variability in speech, such as differences in recording conditions (channel variability) and the same speaker speaking at different times (intersession variability). While techniques like Factor Analysis (FA) and its probabilistic version, Probabilistic Principal Component Analysis (PPCA), have been successful in reducing the dimensionality of speech features and capturing this variability, they are largely unsupervised. This means they don't explicitly use the speaker's identity during the feature extraction process, even though this information is often available during training. The paper addresses this gap by proposing a supervised approach that explicitly incorporates speaker labels to learn more discriminative features.

Methods

The core contribution is the Supervised Probabilistic Principal Component Analysis (SPPCA) model. SPPCA builds upon the Probabilistic Principal Component Analysis (PPCA) framework, which is itself a probabilistic approach to dimensionality reduction.

  • PPCA (Unsupervised): Learns latent variables (factors) that capture the main variance in the acoustic data, effectively reducing feature dimensionality without using speaker labels.
  • SPPCA (Supervised): Extends PPCA by incorporating speaker labels during training. It models both the acoustic data (M) and the speaker identity (S) using a shared set of latent variables (ω). This allows the model to learn a projection that is not only good at reconstructing the acoustic data but also at capturing speaker-specific information. The paper derives an EM algorithm for learning SPPCA parameters.
  • Weighted SPPCA: To bridge the gap between supervised methods like SPPCA (which can be seen as a probabilistic LDA) and the margin-maximizing nature of Support Vector Machines (SVMs), a weighted combination of PPCA and SPPCA is proposed. This allows for a tunable balance between the unsupervised variance maximization of PPCA and the discriminative power of SPPCA.

The learned features from these models are then used as input to an SVM classifier, often after intersession compensation techniques like Weighted Class Covariance Normalization (WCCN), Nuisance Attribute Projection (NAP), and Linear Discriminant Analysis (LDA).

Datasets

  • Training Data: NIST 2004, 2005, 2006 SRE enrollment data, Switchboard II Phase 2 and 3, and Switchboard Cellular Part 1 and 2 were used to train the PPCA and SPPCA models. A gender-dependent UBM was trained on NIST 2004, 2005, 2006 SRE enrollment data. WCCN was trained on the same data as the UBM. LDA and NAP were trained on the same data as PPCA.
  • Evaluation Data: The 5-minute telephone-to-telephone male condition from the NIST 2008 Speaker Recognition Evaluation (SRE) corpus was used for evaluation.
  • SVM Background Impostors: NIST 2004 and 2005 SRE enrollment data were used as SVM background impostors.

Key Insights and Ideas

  • Leveraging Speaker Labels: The fundamental insight is that explicitly using available speaker labels during feature extraction can lead to more discriminative representations than purely unsupervised methods. SPPCA demonstrates this by outperforming PPCA.
  • SPPCA as Probabilistic LDA: The paper notes that SPPCA can be viewed as a probabilistic version of Linear Discriminant Analysis (LDA), suggesting its inherent ability to separate speakers.
  • Bridging Supervised and Margin-Based Classifiers: The Weighted SPPCA model is a clever attempt to reconcile the different optimization criteria of supervised feature extraction (like LDA/SPPCA) and margin-based classifiers like SVMs, potentially leading to better overall performance.
  • Impact of Inter-session Compensation: The effectiveness of SPPCA is shown to be enhanced when combined with established inter-session compensation techniques like WCCN and NAP.

Key Results

  • SPPCA outperforms PPCA: In most tested scenarios, SPPCA achieved lower Equal Error Rates (EER) and minDCF (minimum detection cost function) compared to PPCA. For example, in the SPPCA-NAP-WCCN case, the EER was 6.07% and minDCF was 2.67, compared to PPCA-NAP-WCCN's 6.14% EER and 2.83 minDCF.
  • SPPCA-LDA-WCCN shows best performance: The best overall performance was achieved with SPPCA-LDA-WCCN, yielding an EER of 5.82% and a minDCF of 2.53.
  • Weighted SPPCA shows promise: While not achieving the absolute best performance in the limited evaluation (a=0.2), the weighted SPPCA was better than PPCA and showed potential for optimization by tuning the weighting parameter 'a'.

Limitations and Future Work

  • Computational Cost of SPPCA: The paper acknowledges that the training process for SPPCA is computationally intensive, making it difficult to extensively evaluate the impact of the weighting parameter 'a' in the Weighted SPPCA model.
  • Parameter Optimization: The performance of the Weighted SPPCA is shown to be dependent on the parameter 'a', suggesting that optimizing this parameter is crucial for achieving the best results.
  • Future Work: The authors suggest that optimizing the parameter 'a' for the Weighted SPPCA is a promising direction. They also imply that further exploration of SPPCA's potential in conjunction with various inter-session compensation techniques could yield further improvements.

Summary updated

BibTeX
@inproceedings{lei2010speaker,
  author = {Yun Lei and John H. L. Hansen},
  title = {{Speaker recognition using supervised probabilistic principal component analysis}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure