Introduction
Automatic Speaker Verification (ASV) systems aim to determine if two speech segments belong to the same individual. However, raw similarity scores often suffer from variability due to acoustic conditions, such as background noise and recording hardware. Score normalization is a standard technique used to mitigate these effects. The current state-of-the-art, Adaptive S-norm (AS-norm), utilizes a static "cohort" of impostor speakers for calibration. While effective, AS-norm is limited by its reliance on fixed, pre-calculated representations that do not adapt to specific datasets or provide optimal regularization.
This paper introduces Learnable Impostor Embeddings for Trainable Adaptive Score Normalization (LIE-TAS-norm). By transforming static impostor data into trainable parameters, the proposed method allows for precise score calibration, achieving significant improvements in detection costs and error rates.
Methodology: Trainable Adaptive Score Normalization
The core innovation of TAS-norm is treating the impostor cohort as a set of learnable parameters rather than a static library. This allows the system to fine-tune impostor representations through simulated verification trials.
Learnable Impostor Embeddings (LIEs)
The impostor cohort is initialized from training data and subsequently optimized via gradient descent. This fine-tuning process enables the system to adjust embeddings to better "challenge" the verification model, leading to more robust normalization.
Sub-center Methodology and Loss Functions
To account for intra-speaker variability, the authors employ a sub-center approach where each impostor is represented by multiple embeddings. During normalization, a selection function (e.g., the "min" function) identifies the most appropriate sub-center for comparison.
The training process is guided by a combination of a log-likelihood-ratio ($C_{LLR}$) cost function and an Auxiliary Impostor Classification (AIC) loss. The AIC loss ensures that learnable embeddings maintain their identity as specific speakers during optimization, preventing model divergence.
Experimental Setup
The performance of LIE-TAS-norm was evaluated across three diverse datasets to prove its robustness:
- VoxCeleb1 & 2: The industry standard, primarily English-language celebrity speech.
- CN-Celeb: A large-scale Chinese speaker dataset.
- SdSV 2020 (Task 2): A dataset focused on Persian speaker verification.
Results and Discussion
LIE-TAS-norm consistently outperformed both the baseline (no normalization) and standard AS-norm across all benchmarks.
Key Performance Gains
- VoxCeleb1-O: The method achieved a 4.11% relative improvement in Equal Error Rate (EER) and a 10.62% improvement in minimum Detection Cost Function (minDCF) compared to standard AS-norm.
- Cross-Dataset Robustness: On the Chinese CN-Celeb dataset, EER improved from 6.98% to 6.68%. Similarly, on the Persian SdSV dataset, EER dropped from 1.81% to 1.74%.
Insights
The success of LIE-TAS-norm highlights the benefit of treating score normalization as a trainable component of the ASV back-end. The sub-center approach proved superior to single-center representations, suggesting that capturing the diversity of an impostor's voice is critical. Furthermore, the consistent gains across different languages indicate that the learnable nature of the cohort effectively captures language-specific acoustic nuances without manual tuning.
Conclusion and Future Work
LIE-TAS-norm represents a significant advancement in score normalization by introducing learnability into the impostor cohort. While the system requires an additional training phase, the resulting performance gains in precision and robustness are substantial. Future research will focus on evaluating the method in out-of-domain scenarios and integrating it into a fully end-to-end trainable verification and calibration pipeline.
Summary updated