Abstract
In modern speaker verification, the standard i-vector approach utilizes significant dimensionality reduction, potentially discarding subtle speaker-specific information. This paper investigates the use of high-dimensional "i-supervectors" and introduces a computationally efficient Probabilistic Linear Discriminant Analysis (PLDA) framework to handle the resulting 27,000+ dimensional feature space. By employing matrix partitioning and non-parametric Gaussianized rank-normalization, the authors demonstrate that modeling speaker identity in the original supervector space is highly competitive with state-of-the-art compressed methods, particularly in raw performance scenarios.
Introduction
Speaker verification systems aim to determine if two speech segments originate from the same individual. The prevailing paradigm involves extracting i-vectors—low-dimensional representations (400–600 dimensions) of high-dimensional Gaussian Mixture Model (GMM) supervectors. However, this process involves a "double-compression" bottleneck: first during i-vector extraction and again during PLDA backend modeling. This paper explores the hypothesis that maintaining data in the high-dimensional supervector space throughout the pipeline can preserve critical information that is typically lost during compression.
PLDA Modeling Framework
The core of the proposed verification system is a PLDA backend that performs hypothesis testing on pairs of feature vectors. Whether using standard i-vectors or the proposed i-supervectors, the model evaluates the likelihood of the observed data under two scenarios: the "same speaker" hypothesis ($H_0$) and the "different speakers" hypothesis ($H_1$).
As shown in Figure 6, $H_0$ assumes both observations ($\phi_1, \phi_2$) are generated from a single latent speaker identity ($h_{1,2}$), while $H_1$ assumes they are generated from independent latent identities ($h_1, h_2$).
Methodology: Scaling to High Dimensions
Transitioning from i-vectors to i-supervectors requires overcoming significant computational hurdles. The authors propose several technical solutions:
1. i-supervector Extraction
Unlike standard i-vectors that use a low-rank total variability matrix, i-supervectors are extracted using a diagonal matrix that retains the full dimensionality of the GMM supervector space. In this study, the resulting vectors have 27,648 dimensions.
2. Efficient PLDA via Matrix Partitioning
Modeling tens of thousands of dimensions with PLDA typically involves inverting massive precision matrices. The authors introduce a matrix partitioning strategy that exploits the block-diagonal structure of these matrices, allowing the required linear algebra to be performed in smaller, computationally manageable blocks.
3. Gaussianized Rank-Normalization
Standard normalization techniques like whitening often fail in high-dimensional spaces due to data sparsity. To address this, the authors utilize a non-parametric normalization method that warps each feature dimension into a standard Gaussian distribution based on its rank, making the high-dimensional data more compatible with the Gaussian assumptions of PLDA.
Experimental Evaluation
The proposed methods were evaluated using the NIST SRE08 core task (short2-short3). Training data included Switchboard and previous SRE datasets (04, 05, 06).
Key Results
The study revealed that i-supervectors naturally fit the Gaussian assumptions of PLDA better than i-vectors. In a "raw" state (without normalization), i-supervectors achieved an Equal Error Rate (EER) of 5.26% for male speakers, significantly outperforming the 6.18% EER of raw i-vectors.
However, when full normalization (whitening, length-norm, and s-norm) was applied, i-vectors regained a slight advantage (4.35% vs. 4.59% EER). This suggests that while i-supervectors contain more raw information, i-vectors currently benefit more from established normalization pipelines. Both PLDA-based methods significantly outperformed the Joint Factor Analysis (JFA) baseline across challenging channel conditions.
Conclusion
This research demonstrates that high-dimensional i-supervector modeling is not only computationally feasible but also offers a robust alternative to standard compressed i-vectors. The findings suggest that the Gaussian assumptions inherent in PLDA are more easily satisfied in higher-dimensional spaces. Future work will focus on refining normalization strategies specifically for ultra-high-dimensional feature spaces to further close the performance gap with compressed methods.
Summary updated