Abstract
This paper explores several strategies for Forensic Voice Comparison (FVC), aimed at improving the performance of the Likelihood Ratios (LRs) when using generative Gaussian score-to-LR models. First, different anchoring strategies are proposed to adapt the LR computation process to the specific case while respecting defined propositions. Second, a fully-Bayesian Gaussian model is employed to address the sparsity in training scores often encountered with these anchoring strategies. Experiments using the NIST 2014 i-Vector challenge set-up demonstrate that the fully-Bayesian model significantly outperforms the standard Maximum-Likelihood approach, providing high robustness in data-sparse scenarios.
Introduction
In forensic voice comparison, experts provide a Likelihood Ratio (LR) to quantify the strength of evidence relative to two competing hypotheses: the prosecution's ($H_p$: the suspect is the speaker) and the defense's ($H_d$: someone else is the speaker). Providing reliable LRs is challenging due to acoustic variability—stemming from transmission channels, background noise, and speaker states—and the inherent sparsity of data in criminal cases. Standard statistical models, such as Maximum Likelihood (ML), often perform poorly with limited data, producing extreme and unreliable LR values that can lead to miscarriages of justice.
Methodology
To address these challenges, this research proposes two main strategies: anchoring and a fully-Bayesian framework.
Anchoring Strategies
Anchoring involves selecting training data that mimics the specific conditions of the forensic case. Two strategies were compared:
- Suspect-Anchored (SA): Training scores are generated using any available recordings from the suspect.
- Reference-Anchored (RA): Training scores are specifically anchored to the actual reference recording used in the case, ensuring the model better reflects the specific acoustic conditions of the evidence.
Fully-Bayesian Gaussian Model
Standard systems typically use ML to estimate the mean and variance of score distributions as fixed parameters. In contrast, the fully-Bayesian (BI) approach treats these parameters as random variables and marginalizes over them. This process results in a Student’s t-distribution rather than a Gaussian. The "heavier tails" of the t-distribution naturally account for uncertainty, preventing the model from outputting overconfident, extreme LRs when data is sparse.
Results and Discussion
The study utilized the NIST 2014 i-Vector Challenge dataset, involving over 12 million trials to test robustness.
LLR Moderation
A critical finding is the "self-moderating" nature of the Bayesian approach. As illustrated in the comparison of transformation functions, the Bayesian log-likelihood ratio (LLR) remains significantly more conservative at the extremes of the score distribution compared to the ML-based LLR, which tends toward infinity.
Robustness to Data Sparsity
The performance was measured using $C_{llr}$ (cost of the log-likelihood ratio), where lower values indicate better calibration. The Bayesian model's advantage is most pronounced when the number of training scores is very low. In the Suspect-Anchored scheme, the ML model's performance degrades rapidly as data decreases, reaching a $C_{llr}$ near 12 with only two training scores. Conversely, the Bayesian model remains stable and informative, with $C_{llr}$ values consistently below 1.0 even with a single training score.
Conclusion
The integration of a fully-Bayesian framework with Reference-Anchored data selection provides a robust solution for forensic voice comparison. By accounting for parameter uncertainty through predictive t-distributions, the system avoids the dangerous overconfidence of traditional ML methods in small-data regimes, ensuring more reliable evidence for judicial proceedings.
Summary updated