Abstract
This paper introduces a discriminative decision function scoring method for speaker recognition within the Joint Factor Analysis (JFA) framework. By proposing an approximate form of the decision function, we present a discriminative approach that re-estimates the contribution of individual speech sound units (Gaussian components). This method allows for better classification by exploiting the discriminative power of specific components. Evaluated on the NIST SRE2010 telephone-telephone core condition, the proposed method demonstrates relative improvements of up to 19% and 15% over conventional frame-by-frame and integration scoring strategies, respectively.
Introduction
Speaker verification systems aim to determine if a given speech sample belongs to a target speaker. Joint Factor Analysis (JFA) has been a dominant paradigm, effectively separating speaker-specific variability from channel-induced variability. However, standard JFA scoring typically assumes that all acoustic components contribute equally to the final decision. This paper addresses this limitation by introducing a Discriminative Decision Function (DDF) that weights Gaussian components based on their discriminative utility.
Methodology
The DDF scoring method involves three primary stages: mathematical approximation, component weighting, and discriminative training.
Taylor Series Approximation
The standard log-likelihood ratio (LLR) in JFA is computationally intensive. To facilitate a discriminative framework, the authors approximate the decision function using a first-order Taylor series expansion ($log(x) \approx x - 1$). This simplification transforms the score into a linear combination of occupation probability differences across Gaussian components.
As shown in Figure 3, the approximated scores maintain a strong linear relationship with the traditional decision function scores, validating the use of this simplification for classification tasks.
Discriminative Weighting and SVM Training
Unlike traditional methods that assign uniform weights to all components, the DDF introduces a weight vector $w$. This vector is optimized to emphasize components that are more effective at distinguishing speakers. To learn these weights, a Support Vector Machine (SVM) with a Generalized Linear Discriminant Sequence (GLDS) kernel is employed. The SVM is trained on a development set containing both target and non-target trials, allowing the system to learn the optimal contribution of each sound unit.
Experimental Results
The system was evaluated using the NIST SRE 2010 core condition (telephone-telephone). Training data included NIST SRE 2004-2006 and Switchboard corpora for the Universal Background Model (UBM) and JFA matrices. NIST SRE 2008 was utilized for weight estimation.
The DDF method was compared against frame-by-frame and integration scoring. The results, summarized in the table below, indicate significant performance gains.
For the pooled set, the DDF method achieved an Equal Error Rate (EER) of 4.88%, representing an 18.8% relative improvement over frame-by-frame scoring (5.98%) and a 15.4% improvement over integration scoring (5.77%). Similar trends were observed for the Minimum Decision Cost Function (minDCF) metrics.
Conclusion
The introduction of a discriminative decision function significantly enhances JFA-based speaker verification. By moving beyond the uniform weighting of Gaussian components and employing SVM-based discriminative training, the system better captures speaker-specific nuances. Future work will focus on the robustness of these discriminative weights across diverse acoustic environments and varying utterance lengths.
Summary updated