Age Agnostic Speaker Verification

• Age Agnostic Speaker Verification is the study of ensuring robust speaker verification across different life stages by addressing vocal aging and age-related variability.
• It employs methods such as adversarial learning, mutual information minimization, and domain-specific fusion to decouple age factors from speaker identity features.
• Evaluation protocols using cross-age and longitudinal benchmarks reveal that system performance degrades with increased age mismatch, highlighting fairness and calibration challenges.

Age Agnostic Speaker Verification (AASV) denotes speaker verification research that seeks robustness to age variation, age mismatch, or age-correlated subgroup disparities, so that identity evidence remains stable even when enrollment and test speech come from different life stages or different age populations. In the recent literature, this objective appears under closely related labels such as cross-age speaker verification (CASV), age-invariant speaker representation, inclusive speaker verification, and age-agnostic systems. The field is motivated by a consistent empirical finding: modern deep speaker verification systems achieve strong aggregate performance under standard protocols, yet their scores, thresholds, and error rates can shift materially with age, elapsed time, and age-dependent vocal change (Qin et al., 2022, Singh et al., 2023, Zhang et al., 2024, Zheng et al., 3 Aug 2025).

1. Conceptual scope and relation to speaker verification

AASV is anchored in the standard automated speaker verification pipeline: a system extracts speaker embeddings from enrollment and test utterances, computes a similarity score, and accepts or rejects a same-speaker hypothesis through a decision threshold. The age-specific difficulty is that speaker identity is not perfectly age-invariant. Across the literature, aging is associated with changes in vocal tract and vocal folds, and with alterations in measurable acoustic properties such as pitch, formants, loudness, fricative spectral moments, semitone standard deviation, and other spectral or phonetic properties. The operational consequence is increased intra-speaker variability across time, reduced similarity scores for genuine trials, and elevated false rejection in age-mismatched conditions (Singh et al., 2023, Ehghaghi et al., 2023).

CASV is the most explicit formulation of this problem. It asks whether two recordings belong to the same speaker when the recordings come from different ages of that speaker. This framing makes age a nuisance factor relative to identity, much as language, channel, or noise mismatches are nuisance factors in other SV settings. AASV is broader: it also includes subgroup fairness analyses, threshold adaptation by age, longitudinal robustness over years, and child–adult unification in deployment settings (Qin et al., 2022, Baali et al., 21 Sep 2025).

The topic spans several distinct use cases. In clinical trials, ASV is proposed as a tool to detect duplicate patient participation, where age robustness is part of a broader fairness and integrity requirement. In benchmark research, cross-age protocols stress test models under large age gaps. In child–adult verification, the main concern is the acoustic mismatch between children’s and adults’ speech, which produces severe performance asymmetry when adult-trained systems are applied to children (Ehghaghi et al., 2023, Zheng et al., 3 Aug 2025).

2. Evaluation protocols and benchmark construction

AASV studies use multiple evaluation paradigms, reflecting different notions of age variation. In a clinical setting, one protocol evaluates a pretrained text-independent TitaNet model in a zero-shot setting on an Alzheimer’s Disease Clinical Trial dataset. Embeddings are compared using cosine similarity, a threshold $\theta$ is tuned per subgroup, and performance is evaluated using Equal Error Rate (EER). The decision rule is: if cosine similarity $>\theta$, classify as same speaker; if cosine similarity $<\theta$, classify as different speakers. Positive-pair and negative-pair counts are defined respectively as

$\sum_{i=1}^m {n_i \choose 2}$

and

$\sum_{i=1}^m {n_i * (N-n_i)/2},$

where $m$ is the number of speakers in the group, $n_i$ is the number of recordings for speaker $i$, and $N$ is the total number of audio files in the group. To study age robustness, speakers are split into Age $\leq 70$ and Age $>\theta$0, with 70 chosen because it is approximately the median and mean age of the dataset (Ehghaghi et al., 2023).

Benchmark-oriented CASV work uses deliberately age-stressed trial construction. Vox-CA is built from VoxCeleb by estimating apparent age from associated face images, then constructing positive trials with large age gaps and negative pairs constrained by nationality and gender. Four test sets are defined: Vox-CA5, Vox-CA10, Vox-CA15, and Vox-CA20, where positive pairs have age gaps of at least 5, 10, 15, and 20 years, respectively. “Only-CA” variants isolate cross-age effects more directly, while the full Vox-CA protocols are harder because they align negative-pair construction with Vox-H cases (Qin et al., 2022).

Longitudinal evaluation extends this logic from discrete age-gap protocols to continuous temporal drift. VoxAging contains 293 speakers, 2,629,100 segments, 7,522 hours of audio-visual data, weekly recording intervals, and up to 17 years of span. It supports two broad settings: an X-Independent setting with VoxAging-EN and VoxAging-ZH divided into time spans, and an X-Dependent setting with age-group analysis and gender analysis. The age-group analysis uses five groups: \<30, 30~40, 40~50, 50~60, and \>60 (Ai et al., 27 May 2025).

A broader robustness benchmark, SVeritas, treats speaker age as one of several real-world stressors. Its age evaluation is subgroup-based rather than explicitly continuous: verification pairs are generated after splitting by gender and then by age, and EER is computed independently for each subgroup. On EARS the age bins are 18–25, 26–35, 36–45, 46–55, 56–65, and 66–75; on Common Voice the bins are Teens, Twenties, Thirties, Forties, Fifties, Sixties, and Seventy+ (Baali et al., 21 Sep 2025).

3. Empirical evidence that current systems are not fully age-agnostic

Across clinical, benchmark, and longitudinal studies, the central result is consistent: speaker verification performance deteriorates as age difference or aging-related drift increases (Ehghaghi et al., 2023, Qin et al., 2022, Ai et al., 27 May 2025, Singh et al., 2023).

Evaluation setting	Comparison	Reported EER
AD clinical-trial ASV	Age $>\theta$1 vs Age $>\theta$2	3.62% vs 4.20%
VoxCeleb-derived CASV	Vox-H vs Vox-CA20	1.939% vs 10.419%
VoxAging-EN, ERes2Net-large	0 years vs 10 years	2.89% vs 3.87%
LCFSH male, long-term ageing	0 vs 20 vs 40 years	5.12% vs 18.68% vs 51.61%

The clinical result is notable because it arises in a deployment-motivated setting rather than a generic speech benchmark. On the Alzheimer’s Disease Clinical Trial dataset, age $>\theta$3 yields an EER of 4.20%, compared with 3.62% for Age $>\theta$4, a difference of 0.58 EER points. The same study reports an overall baseline EER of 3.10%, indicating that subgroup performance can be meaningfully worse than the aggregate result (Ehghaghi et al., 2023).

The Vox-CA benchmark shows that the problem becomes much harder as age gap grows. A baseline ResNet34-GSP-ArcFace system degrades from 0.962% EER on Vox-O and 1.939% on Vox-H to 10.419% on Vox-CA20. The “only-CA” conditions also worsen monotonically with age gap, from 1.953% on only-CA5 to 8.185% on only-CA20. These results isolate age as a major source of failure even before adding nationality and gender constraints (Qin et al., 2022).

Short-term and long-term aging studies using ECAPA-TDNN provide a complementary score-level view. On VoxCeleb1-Age-Enriched, EER generally worsens with age difference for both male and female English speakers, although not perfectly monotonically at every single year. On LCFSH, the long-term effect is much stronger: male EER rises from 5.12% at 0 years to 18.68% at 20 years and 51.61% at 40 years, while female EER rises from 10.04% to 21.67% and 40.61% (Singh et al., 2023).

VoxAging confirms that deterioration persists in dense, weekly longitudinal data and across several advanced SV systems. For English, ERes2Net-large changes from 2.89% EER at 0 years to 3.87% at 10 years, while ECAPA-TDNN changes from 4.07% to 5.36%. For Mandarin, all models have higher EERs and generally larger degradation than in English. The paper also reports that similarity scores decline steadily over time, with the average score in English falling below 0.5 after about 500 weeks and in Mandarin after about 400 weeks (Ai et al., 27 May 2025).

4. Modeling strategies for age robustness

One family of methods seeks age-invariant embeddings by explicit factorization of age and identity. In "Cross-Age Speaker Verification: Learning Age-Invariant Speaker Embeddings" (Qin et al., 2022), the feature embedding is modeled as

$>\theta$5

with an Age-Related Extractor producing

$>\theta$6

The identity branch is supervised by

$>\theta$7

the age branch by

$>\theta$8

and age leakage in the identity branch is reduced with a Gradient Reversal Layer through

$>\theta$9

This age-decoupling adversarial learning framework, ADAL, improves Vox-CA performance over the baseline, with the largest gains at the largest age gap.

A second line of work replaces adversarial age confusion with mutual information minimization. "Disentangling Age and Identity with a Mutual Information Minimization Approach for Cross-Age Speaker Verification" (Zhang et al., 2024) decomposes the initial embedding as

$<\theta$0

and defines mutual information between the age and identity embeddings by

$<\theta$1

Using a CLUB-based estimator, the method minimizes dependence between $<\theta$2 and $<\theta$3 and introduces an aging-aware weight

$<\theta$4

to focus more on large age gaps. The final backbone objective is

$<\theta$5

Compared with ResNet34+ADAL, this method yields 1.53% relative improvement in EER and 4.79% relative improvement in minDCF on Vox-CA test sets.

A third strategy addresses the child–adult domain gap through domain-specific encoders and soft domain-weighted fusion. "An Age-Agnostic System for Robust Speaker Verification" (Zheng et al., 3 Aug 2025) combines an adult ECAPA-TDNN model, a child-finetuned ECAPA-TDNN model, and a domain classifier whose softmax output is $<\theta$6. The final representation is

$<\theta$7

This formulation avoids hard routing between child and adult subsystems. In the reported experiments, AASV-Large matches C-SV-Large almost exactly across most child age groups while remaining far closer to adult performance than C-SV, and AASV-Small reduces the adult-domain penalty dramatically relative to child fine-tuning alone.

Threshold adaptation constitutes a fourth line of work. One paper proposes a context adaptive thresholding framework for SV, where the context can be gender or age, and reports that by using an age-specific threshold it can significantly reduce FRR for certain age groups for desired FAR (Jain et al., 2021). This approach does not enforce age-invariant embeddings; instead, it calibrates decision thresholds to compensate for subgroup-dependent score distributions.

5. Clinical, fairness, and demographic implications

In healthcare and trial integrity settings, age robustness is coupled to fairness and operational risk. The Alzheimer’s Disease Clinical Trial study uses a proprietary dataset with 659 speakers and 7,084 audio recordings collected longitudinally over 48 weeks through picture description, phonemic verbal fluency, and semantic verbal fluency. The task is duplicate patient participation detection, and the study reports that ASV performance is slightly better on male speakers than on female speakers, degrades for individuals who are above 70 years old, is comparatively better for non-native English speakers than for native English speakers, is negatively affected by clinician interference, noisy background, and unclear participant speech, and tends to decrease with an increase in the severity level of AD (Ehghaghi et al., 2023).

These findings are explicitly framed as fairness concerns. The study states that voice biometrics raise fairness concerns as certain subgroups exhibit different ASV performances owing to their inherent voice characteristics. For age specifically, the system is interpreted as not fully age-agnostic, because age-related voice differences may be mistakenly treated as speaker identity cues. A plausible implication is that age and pathology can interact in practical verification settings, even when age is not the only source of variability. The same paper notes that if older adults are more likely to speak less clearly or in noisier conditions, part of the age effect could be indirectly mediated by recording quality, though the paper does not explicitly claim that (Ehghaghi et al., 2023).

SVeritas generalizes this fairness perspective beyond a single clinical corpus. It identifies age as a first-class robustness dimension and reports that models suffer substantial performance degradation in scenarios involving age mismatches. Its subgroup analyses show that older male and female speakers (60+) consistently underperform compared to younger groups, and the authors interpret these disparities as evidence that demographic imbalance in training data may contribute to uneven generalization and reduced fairness, particularly in age and ethnicity subgroups with limited representation (Baali et al., 21 Sep 2025).

The child–adult setting raises a related but distinct inclusivity issue. Adult-trained systems perform strongly on VoxCeleb adults but degrade sharply on OGI children, while children-finetuned systems improve on child speech but often forget adult-domain knowledge. The explicit aim of AASV in this literature is therefore not merely adaptation to children, but a unified representation that works across both domains (Zheng et al., 3 Aug 2025).

6. Limitations, unresolved issues, and research directions

Current AASV research is constrained by data, label quality, and problem formulation. Several studies use coarse age bins rather than continuous age modeling. The clinical-trial study tests only two bins, $<\theta$8 and $<\theta$9, over an age range limited to 55–80, and the authors note that this design cannot identify a more precise age transition point beyond the coarse 70-year threshold (Ehghaghi et al., 2023). Vox-CA relies on estimated age labels derived from face age estimation rather than ground-truth age, which the authors regard as sufficient to rank recordings chronologically but not as perfectly accurate labels (Qin et al., 2022). The mutual-information method also depends on age labels and acknowledges that the age groups are coarse (Zhang et al., 2024).

Generalizability remains uncertain. Some results are language-dependent and even reverse in gender direction across corpora. In English VoxCeleb, ageing affects female speakers to a greater degree than male speakers, while in Finnish LCFSH it has a greater impact on male speakers than female speakers (Singh et al., 2023). VoxAging reports that the Mandarin trend is consistent with English but more severe overall, and its Mandarin temporal coverage is shorter than English (Ai et al., 27 May 2025). These results indicate that age robustness cannot be reduced to a single universal correction.

The literature also separates infrastructural contributions from algorithmic ones. VoxAging is explicitly a dataset and analysis paper rather than a new AASV algorithm paper, yet its dense weekly sampling, multi-year span, and bilingual design provide an enabling benchmark for template updating strategies, score calibration under age mismatch, and training age-robust embeddings (Ai et al., 27 May 2025). By contrast, ADAL, MI minimization, and domain-weighted child–adult fusion are mitigation strategies, but none fully eliminates the cross-age gap. Even after improvement, cross-age EER remains much worse than standard VoxCeleb verification on hard protocols (Qin et al., 2022, Zhang et al., 2024).

A recurring research direction in the cited work is therefore explicit age-aware evaluation rather than reliance on aggregate SV accuracy alone. The papers repeatedly suggest age-balanced training, calibration, or domain adaptation; avoidance of age-correlated voice features as identity cues; time-aware evaluation; and fairness checks by age group, especially in healthcare and inclusive deployment settings (Ehghaghi et al., 2023, Ai et al., 27 May 2025, Baali et al., 21 Sep 2025). Taken together, the literature indicates that current systems are not fully age-agnostic, but it also establishes the benchmark infrastructure, modeling strategies, and subgroup analyses required to make age robustness a tractable and measurable objective.