Abstract
Traditional speaker verification systems often rely on proxy loss functions that do not directly optimize for the final evaluation metric, the Detection Cost Function (DCF). This paper presents a novel differentiable loss function, aDCF, which incorporates the decision threshold as a learnable parameter. Experimental results on the RSR2015-PartI corpus demonstrate that this approach significantly improves system calibration and error rates, achieving a 24% relative improvement over standard Cross-Entropy baselines.
Introduction
Speaker Verification (SV) systems are tasked with confirming a speaker's identity based on a voice sample. This process inherently involves a trade-off between two error types: False Acceptance Rate (FAR) and False Rejection Rate (FRR). In industrial applications, performance is typically measured using the Detection Cost Function (DCF), which weights these errors according to the specific security or convenience requirements of the application.
A persistent challenge in deep learning-based SV is the disconnect between training objectives and evaluation metrics. Common loss functions like Cross-Entropy or Triplet Loss optimize for classification accuracy or embedding distance but ignore the operational decision threshold. This paper addresses this gap by proposing a differentiable approximation of the DCF.
Methodology
The core contribution is the aDCF (approximated DCF) loss function. To make the non-differentiable DCF compatible with backpropagation, the authors utilize sigmoid functions to approximate the discrete counting of FAR and FRR. Furthermore, the decision threshold ($\\Omega$) is treated as a trainable parameter, allowing the network to learn the optimal boundary between "Accept" and "Reject" decisions during training.
System Architecture
The researchers evaluated two distinct architectural configurations, as shown in Figure 1. Type A represents a direct end-to-end model, while Type B incorporates a trainable back-end initialized from a pre-trained Type A model.
A critical feature of both architectures is the use of a frame-to-state alignment mechanism to generate a "Supervector." Unlike global average pooling, which can lose temporal nuances, this mechanism preserves phonetic structure. This is particularly important for text-dependent SV, where the phonetic content of the utterance is fixed and informative.
Experimental Results
The proposed methods were tested on the RSR2015-PartI dataset using 20-dimensional MFCC features. The performance was evaluated using Detection Error Tradeoff (DET) curves, which plot FRR against FAR.
As shown in Figure 2, the aDCF loss (green curve) consistently yields lower error rates across various operating points compared to Cross-Entropy (CE), Triplet Loss (TrLoss), and aAUC loss. Notably, the aDCF system achieved an Equal Error Rate (EER) of 0.56% when combined with S-norm score normalization.
Key Insights
- Threshold Awareness: By making the threshold learnable, the model achieves better intrinsic calibration, reducing the need for complex post-processing.
- Robustness to Normalization: aDCF maintains high performance even when score normalization is applied, whereas other losses like Triplet Loss show significant degradation in EER (from 0.87% to 1.05%).
- Low FAR Optimization: The aDCF loss is particularly effective at low FAR regions, which is a primary requirement for high-security commercial applications.
Conclusion
The aDCF loss function provides a robust framework for optimizing speaker verification systems directly for their operational goals. By bridging the gap between training and evaluation, the proposed system achieves state-of-the-art performance on text-dependent tasks. Future research will explore the impact of varying the FAR/FRR weights ($\gamma$ and $\beta$) to tailor systems for specific deployment scenarios.
Summary updated