Paper deep dive
InterBias-SV: Compound Conditions in Speaker Verification
Kamel Kamel, Hridoy Sankar Dutta, Keshav Sood, Sunil Aryal
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Speaker verification systems encounter combinations of noise, channel distortion, and changes in speech. Evaluating each condition separately does not establish whether their effects add. InterBias-SV organises this question around a four-term comparison: joint error, two marginal errors, and a common reference. Its results artefact contains 4,068 scored records across 17 experiments, 12 encoder labels, and six speech corpora, totalling 12 million trial evaluations. Three experiment families contain the same-corpus terms needed to compute additive contrasts. For labels assigned to speaker-trained encoders, their mean contrasts are +0.0026, +0.0088, and +0.0024 in equal error rate (EER), with larger variation across settings. These descriptive averages do not establish equivalence to additivity: trial matching, checkpoint identity, and parts of the condition metadata remain unverified. We also examine two interpretation problems. Near-chance EER can make additive predictions difficult to interpret, but chance performance is not a hard EER ceiling, and correlation with the prediction does not identify a saturation mechanism. Ratios of demographic gaps are unstable when their clean reference is near zero; absolute gaps provide a more direct summary. The benchmark provides condition definitions, analysis scripts, and explicit requirements for interpretable compound-condition comparisons, while separating recomputable summaries from claims that require further experimental validation.
Tags
Links
- Source: https://arxiv.org/abs/2609.36500v1
- Canonical: https://arxiv.org/abs/2609.36500v1
Trouble viewing inline? Open PDF directly →
Full Text
90,308 characters extracted from source content.
Expand or collapse full text
InterBias-SV: Compound Conditions in Speaker Verification Kamel Kamel Hridoy Sankar Dutta Keshav Sood Sunil Aryal Affiliation: School of Information Technology, Deakin University Affiliation: Waurn Ponds, VIC, Australia Email: kamel.kamel@research.deakin.edu.au Abstract Speaker verification systems encounter combinations of noise, channel distortion, and changes in speech. Evaluating each condition separately does not establish whether their effects add. InterBias-SV organises this question around a four-term comparison: joint error, two marginal errors, and a common reference. Its results artefact contains 4,068 scored records across 17 experiments, 12 encoder labels, and six speech corpora, totalling 1.20×1071.20× 10^7 trial evaluations. Three experiment families contain the same-corpus termsneeded to compute additive contrasts. For labels assigned to speaker-trained encoders, their mean contrasts are +0.0026+0.0026, +0.0088+0.0088, and +0.0024+0.0024 in equal error rate (EER), with larger variation across settings. These descriptive averages do not establish equivalence to additivity: trial matching, checkpoint identity, and parts of the condition metadata remain unverified. We also examine two interpretation problems. Near-chance EER can make additive predictions difficult to interpret, but chance performance is not a hard EER ceiling, and correlation with the prediction does not identify a saturation mechanism. Ratios of demographic gaps are unstable when their clean reference is near zero; absolute gaps provide a more direct summary. The benchmark provides condition definitions, analysis scripts, and explicit requirements for interpretable compound-condition comparisons, while separating recomputable summaries from claims that require further experimental validation. 1 Introduction A speaker verification (SV) system may receive noisy, compressed speech from a caller whose speaking style differs from enrolment. A benchmark can measure the error associated with each condition, but those measurements alone do not show how the conditions combine. The practical question is whether a joint condition produces the error predicted by adding the two single-condition changes to a common reference. Answering it requires measuring the joint condition and its reference terms under comparable sampling and scoring procedures. Existing resources establish useful but different comparisons. VoxCeleb and the VoxCeleb Speaker Recognition Challenges evaluate recognition in unconstrained recordings (Nagrani et al., 2017; Chung et al., 2018; Huh et al., 2024). SVeritas evaluates a broad set of stressors, including compounded noise, reverberation, and codec conditions (Baali et al., 2025). Such joint-condition results describe difficulty; attributing a departure from additivity additionally requires the corresponding marginals and an explicit contrast. Our focus is this comparison, rather than a claim that prior benchmarks evaluate only isolated stressors. InterBias-SV combines condition definitions, an evaluation harness, and an analysis of an existing results artefact. Its proposed evaluation unit is a four-term comparison for one encoder and a specified factor setting. We distinguish acoustic manipulations from population strata: an age-group contrast measures how codec effects differ between populations, not the effect of changing a speaker’s age. We also distinguish the presence of four numerical terms from evidence that they form a controlled experiment. That distinction matters here: the artefact permits three families of same-corpus contrasts, but the supplied code and metadata do not establish all the matching and provenance requirements. The contributions are: 1. A benchmark organisation for compound-condition SV evaluation, with 17 recorded experiments, 4,068 scored records, and scripts that reconstruct its tabulated summaries (Section 4). 2. Explicit requirements for interpreting additive contrasts: comparable populations and trial sampling, measured reference terms, and a clearly defined metric scale (Section 3). The accompanying audit identifies which requirements the current artefact does and does not establish. 3. Descriptive evidence on compound-condition errors and demographic gaps (Section 5). Small positive mean contrasts coexist with large setting-level deviations. Near-chance performance and small reference gaps limit interpretations of additive and ratio summaries. The paper contributes an evaluation framework and empirical analysis, rather than a new encoder or a newly collected speech dataset. Its usefulness depends on making both the measurements and their limits reproducible. 2 Related Work and Positioning Robustness evaluation. VoxCeleb provides in-the-wild speech and verification protocols (Nagrani et al., 2017; Chung et al., 2018); VoxSRC extends evaluation across training and adaptation settings (Huh et al., 2024). Noise and reverberation augmentation are also established components of speaker-embedding pipelines (Ko et al., 2017; Snyder et al., 2018). SVeritas is the closest benchmark: it evaluates duration, speaking conditions, noise, reverberation, channels, codecs, age, spoofing, and adversarial perturbations, with demographic breakdowns (Baali et al., 2025). It includes compound conditions. InterBias-SV’s specific emphasis is the contrast between a joint error and the additive prediction from its own reference terms. This is an application of a factorial contrast, not a new statistical definition of interaction. Representations and demographic evaluation. SUPERB evaluates speech representations across downstream tasks (Yang et al., 2021). WavLM, HuBERT, and UniSpeech-SAT support speaker-related applications (Chen et al., 2022a; Hsu et al., 2021; Chen et al., 2022b), but a frozen mean-pooled representation is a different system from one with a trained speaker head. Our comparison retains that distinction. Work on demographic bias in speech recognition (Koenecke et al., 2020) and speaker recognition (Hutiri and Ding, 2022; Fenu et al., 2021; Peri et al., 2023) motivates subgroup reporting. Here we examine how reported gaps vary across conditions, without treating subgroup EERs as a fairness assessment at a shared deployment threshold. Table 1: Evaluation emphasis of related resources. The comparison describes reported analyses, not an exhaustive claim about what their data could support. Resource Evaluation emphasis VoxCeleb / VoxSRC In-the-wild verification and challenge protocols SUPERB Downstream evaluation of speech representations Fair Voice Biometrics Demographic imbalance and group fairness SVeritas Broad stressor coverage, including compound conditions InterBias-SV Additive contrasts and their sampling and metric requirements 3 Task, Estimand and Validity Conditions Task and scoring. The intended task is text-independent SV. A frozen encoder maps an utterance to an L2L_2-normalised embedding, and cosine similarity scores each genuine or impostor trial. The artefact reports equal error rate (EER), normalised minimum detection cost (minDCFminDCF; target prior 0.010.01, miss and false-accept costs both 1), and area under the receiver operating characteristic curve (AUC). The supplied EER routine selects the ROC point with the smallest absolute difference between false-positive and false-negative rates and averages those rates; it does not interpolate the crossing. All EERs here are fractions, so 0.010.01 corresponds to one percentage point. No score calibration, score normalisation, fine-tuning, or domain adaptation is specified. A common scoring rule removes one source of variation, but does not by itself control trial sampling or establish checkpoint identity. The estimand. For conditions A and B and reference cleanclean, define EER^(A∩B)=EER(A)+EER(B)−EER(clean),IE(A,B)=EER(A∩B)−EER^(A∩B). EER(A∩ B)=EER(A)+EER(B)-EER(clean), (A,B)=EER(A∩ B)- EER(A∩ B). (1) The non-linearity index is NLI(A,B)=EER(A∩B)/EER^(A∩B)NLI(A,B)=EER(A∩ B)/ EER(A∩ B) when the denominator exceeds 10−910^-9; zero or negative predictions retain their IEIE values but have undefined NLINLI. Small positive denominators can still make the ratio unstable. Positive IEIE indicates error above the additive prediction on the EER scale; negative IEIE indicates error below it. Neither sign alone identifies a causal mechanism. Additivity is scale-dependent and does not imply statistical independence of the factors. We report means and standard deviations (SDs) over encoder–setting combinations; these SDs are not confidence intervals. Requirements for interpretation. C1 (comparable source population): all terms should refer to the same corpus and a documented target population. A shared corpus alone does not control speaker, session, or content composition. C2 (comparable trials): retain trial identities across acoustic transformations, and match speaker and content composition when a factor requires different recordings. Report sampling, class counts, exclusions, and uncertainty. Independent trial draws can estimate population contrasts under an appropriate design, but their variation must be accounted for. C3 (measured terms): both marginals and the reference must be measured; substituting the reference for a missing marginal changes the estimand. C4 (metric domain): check the attainable metric range and the stability of ratio denominators. These are requirements for the intended analysis, not conditions certified by the presence of result rows. Three experiment families have the same-corpus terms needed for numerical recomputation: E13 (age group × codec), E14 (emotion × noise), and E17 (recorded drift label × codec). The updated E13 and E14 exports identify a common protocol cohort within each experiment and include checkpoint hashes. They are labelled paired-age and matched-RAVDESS runs, respectively; the aggregate CSV itself contains no trial identities, so C2 cannot be independently checked from this release. E13 compares an age stratum with the overlapping all-age population, rather than measuring within-speaker ageing. E14 uses all seven non-neutral RAVDESS emotions crossed with three noise labels and three SNRs. E17 still requires recording-level dates that the bundled legacy loader does not provide (Section 7). E15 and E18 are reported as within-corpus clean/codec summaries, not isolated pathology or accent interactions: their reference populations belong to different corpora. E16 (gender × noise) is analysed through subgroup gaps. E20’s recorded condition values are retained descriptively. The run audit found measured reference terms, but temporal and age proxies limit their interpretation. The analysis therefore distinguishes recomputable four-term contrasts from verified controlled experiments. Reference and two marginalsEER(clean),EER(A),EER(B)EER(clean),\ EER(A),\ EER(B)Additive predictionEER^(A∩B) EER(A∩ B)Measured joint errorEER(A∩B)EER(A∩ B)IE,NLIIE,\ NLICheck population, trials,reference terms, and metric scale Figure 1: Four terms define an additive comparison. Their numerical availability permits recomputation; comparable sampling and verified condition provenance are needed to interpret the contrast. Chance performance and mathematical coupling. With fixed score orientation, EER lies in [0,1][0,1]; 0.50.5 is the chance reference, not a hard upper bound. The supplied routine neither reverses nor clips scores, and the artefact contains EERs above 0.50.5. An additive prediction above chance therefore does not force a negative interaction. More generally, if an error metric is bounded above by U, then IE≤U−EER^IE≤ U- EER; a prediction above U necessarily gives a negative contrast. For this EER implementation, U=1U=1. Correlation between IEIE and EER EER is also not an independent test of saturation: the latter is subtracted in the former. Writing Y=EER(A∩B)Y=EER(A∩ B) and X=EER^(A∩B)X= EER(A∩ B) gives Cov(Y−X,X)=Cov(Y,X)−Var(X)Cov(Y-X,X)=Cov(Y,X)-Var(X). A negative correlation can consequently arise without a ceiling mechanism. We report the additive range and this correlation as descriptive checks, without using them to certify or exclude an encoder family. 4 Benchmark Construction Corpora and encoder configurations. The records name RAVDESS (Livingstone and Russo, 2018), TORGO (Rudzicz et al., 2012), EARS (Richter et al., 2024), Common Voice (Ardila et al., 2020), L2-ARCTIC (Zhao et al., 2018), and VoxCeleb1 (Nagrani et al., 2017). RIRS_NOISES is the configured external-noise source (Ko et al., 2017). The audio front end resamples to 16 kHz, rejects clips shorter than 1 s, and truncates those longer than 10 s. The artefact includes speaker counts for some records, but does not provide complete corpus-release, audio-duration, or sampling manifests. Its trial total sums scoring operations across conditions and models, not distinct utterances or independent speaker pairs. Twelve model labels appear. Seven are assigned to speaker encoders: X-Vector (Snyder et al., 2018), ECAPA-TDNN and ECAPA-TDNN-Large (Desplanques et al., 2020), RawNet3 (Jung et al., 2022), CAM++ (Wang et al., 2023b), TitaNet (Koluguri et al., 2022), and ReDimNet (Yakovlev et al., 2024). The wrappers use SpeechBrain (Ravanelli et al., 2021), WeSpeaker (Wang et al., 2023a), or model-specific loaders, with unrecorded fallback paths. Thus these labels do not verify seven distinct resolved speaker checkpoints. Five self-supervised learning (SSL) configurations use WavLM-Base, WavLM-Base+, WavLM-Large (Chen et al., 2022a), UniSpeech-SAT (Chen et al., 2022b), and HuBERT-Large (Hsu et al., 2021). Their intended utterance representation is the temporal mean of the final hidden layer, without a trained speaker head. We retain the recorded family split for summarisation; it is not a controlled comparison of training paradigms. Original model-label spellings are retained in tables to permit exact CSV lookup. Operators and provenance. The harness specifies telephony encode–decode operations with ffmpeg, noise at 5, 15, and 25 dB signal-to-noise ratio (SNR), 20 ms packet-loss masks, and time-scale perturbations. The run audit identifies E04 as recorded EARS styles and E05 as SBC, AAC, aptX and aptX-HD encode–decode conditions. The DSP/AAC proxy paths in the bundled legacy harness do not describe those retained runs. Codec round trips do not model a wireless radio channel. Other transformations also need verification. The environmental and crosstalk operators draw from the same noise directory, and revert to Gaussian noise when files are unavailable. Codec failures can return unchanged audio. Synthesis and conversion wrappers name VoiceCraft (Peng et al., 2024), E2 TTS (Eskimez et al., 2024), kNN-VC (Baas et al., 2023), and FreeVC (Li et al., 2023), but also return original audio on failures. The updated export includes augmentation and decode-failure counts for some records, but these fields do not cover every experiment. Appendix A.1 separates configured operations from verified execution. Trials, splits, and leakage. The general sampler requests balanced genuine and impostor trials with a cap of 10410^4, using a module-level random generator initialised with seed 42. Repeated calls advance that generator; a fixed initial seed does not give the same trial list to every model or condition. Scoring failures can change class balance. The general genuine-pair sampler uses different recordings, but this constraint is not enforced by every specialised sampler. Audio augmentation randomness and cache state are also not fully captured in the result rows. No training or validation split is defined for benchmark-specific learning; the intended protocol freezes pre-trained encoders. This does not establish that evaluation speakers were unseen in pre-training. Several checkpoints use VoxCeleb, and E07, E17, and E20 reference VoxCeleb1. Speaker-level overlap has not been audited. Future model selection on these public conditions would also require a separate held-out protocol to assess generalisation. Coverage and availability. The submission CSV contains 4,068 scored records from 17 experiments and 12 model labels, totalling 11,975,026 trial evaluations. It includes ten single-factor, six two-factor, and one nominal three-factor experiment. E13 and E14 use only the replacement cohorts, each covering 12 models; the single-model legacy E13 cohort is excluded. E21 is removed because the updated cohort covers only one of 13 model labels, leaving the rest as legacy results. E09, E12, and E19 remain absent. E15 and E18 are restored from the historical export as descriptive results, with 240 and 120 scored rows respectively. Their corpus-level counts and aggregate metrics agree with the available run audit; their cross-corpus contrasts are excluded from the interaction analysis. The retained roster contains the same 12 models in every experiment, with complete coverage of each experiment’s observed condition grid. ERes2Net is excluded because it lacks four of the retained experiments. No missing measurements are imputed. Appendix A.2 reports coverage. E05, E10, and E11 contain operator labels in the dataset field rather than identifiable corpus names; these are shown as unrecorded. The anonymised supplement contains the selected results, analysis scripts, and available legacy harness. It permits recomputation of aggregate summaries, but the harness is not established as the implementation of the updated runs. No corpus audio is included. Corpus acquisition and reuse remain subject to original distributor terms; an artefact licence and permanent repository have not yet been specified. Intended use. The intended uses are condition-level comparisons, additive contrasts with verified reference terms, and stress-conditioned subgroup analysis. A low interaction effect is not itself a robustness score: a system can have low accuracy under every condition and still have a small contrast. Repeated tuning on public cells can also overfit the benchmark. Absolute error, balanced coverage, uncertainty, and held-out conditions are needed alongside interaction summaries. 5 Results Encoder summaries. Macro-averaging first averages condition values within experiments and then weights experiments equally. CAM++ has macro EER 0.098, ECAPA-TDNN 0.101, and ECAPA-TDNN-Large 0.102; nominal speaker-trained labels span 0.098–0.181. Mean-pooled SSL labels span 0.331–0.415 (Figure 2; Appendix B). All 12 labels cover the same 17 experiments and each observed condition grid is complete. The macro ordering remains descriptive: common coverage alone does not establish matched trials or comparable sampling across encoders. Checkpoint hashes are present for a subset of the results; sampling and provenance still limit comparisons. Figure 2: Left: macro EER by recorded model label; all 12 encoders cover the same 17 experiments. Right: E01 on RAVDESS, averaging available encoders for each condition. Bars show between-encoder SD, not uncertainty in a mean. Both panels use EER fractions. Family assignments and condition labels follow the artefact. Emotion and enrolment–test mismatch. On RAVDESS, the neutral/neutral mean is 0.083 EER and the neutral→ mean is 0.256, a ratio of 3.11×3.11×. Matched fearful speech has mean EER 0.191 (2.32×2.32× the reference). Each neutral-to-emotion mismatch has a higher mean than the corresponding same-emotion condition (Appendix C). This pattern is consistent with a contribution from emotional mismatch, but does not isolate a causal mechanism or rank emotion against stressors on other corpora. All 15 conditions cover the same 12 model labels; matched and mismatched trials have different budgets. Between-encoder SDs are 0.11–0.16, not confidence intervals for the condition differences. The updated E01 export contains RAVDESS only. Same-corpus additive contrasts. For E13, E14, and E17 we reconstruct Equation (1) from the CSV, retaining settings with all four finite terms for the same model label. This verifies the arithmetic, not the sampling or condition provenance (Section 3). Table 2 retains all such settings without a post-hoc family-level validity filter. Table 2: Descriptive additive contrasts from available same-corpus terms. Means and SDs are across model–setting combinations; n counts those combinations, not independent replicates. EER uses fractional units; “IE>0IE>0” is a percentage. SV denotes nominal speaker-trained labels and SSL denotes frozen mean-pooled configurations. E13 compares an age stratum with an overlapping all-age reference; E17’s temporal provenance is unresolved. For E14 SV, NLINLI averages 198 positive-denominator cells; n=441n=441 includes all contrasts, including 243 with nonpositive predictions. SV labels SSL labels Experiment IEIE SD NLINLI IE>0IE>0 n IEIE SD NLINLI IE>0IE>0 n E13 age × codec +0.0026 0.0204 1.089 52 84 -0.0063 0.0209 0.973 35 60 E14 emotion × noise +0.0088 0.0301 1.549 34 441 -0.0465 0.0540 0.925 22 315 E17 drift × codec +0.0024 0.0094 1.137 62 84 +0.0049 0.0198 1.014 57 60 The nominal speaker-trained group has mean contrasts +0.0026+0.0026, +0.0088+0.0088, and +0.0024+0.0024 for E13, E14, and E17; 52%, 34%, and 62% of combinations have positive signs. Mean NLIs are 1.089, 1.549, and 1.137. These average per-cell ratios over positive predictions only; in E14, 243 of the 441 SV predictions are nonpositive and have no NLI, while their contrasts remain in the IE summary. Small denominators make NLI particularly unstable in this experiment. Signed mean departures are small, but SDs of 0.009–0.030 exceed them. Individual E14 SV contrasts range from −0.059-0.059 to +0.139+0.139. Cancellation and zero contrasts can therefore obscure substantial setting-level deviations. The table does not establish equivalence to additivity or justify replacing joint-condition tests with marginal measurements. Appendices E.1 and E.2 report this variation. Near-chance predictions. For mean-pooled SSL labels in E14, the mean contrast is −0.0465-0.0465, and 22% of combinations have positive contrasts. The additive prediction averages 0.438 and reaches 0.670; its correlation with IEIE is r=−0.706r=-0.706 (Table 3, Figure 3). For nominal speaker-trained labels, the mean prediction is 0.039 and r=+0.504r=+0.504. These differences describe the supplied configurations, without identifying a saturation mechanism: EER can exceed chance and the correlation shares a term with the contrast. E13’s SSL predictions remain below 0.5, so a chance-ceiling criterion would not exclude that family either. Figure 3: Left: mean additive contrast by experiment and recorded family; bars show SD across model–setting combinations. Right: E14 contrasts versus additive predictions, with descriptive least-squares fits. The dotted line marks chance EER, not an upper bound. The shared prediction term contributes to the negative correlation. Table 3: Additive-prediction summaries. EER^¯ EER and maxEER EER are the mean and maximum prediction; r is Pearson correlation between the prediction and IEIE. These are descriptive diagnostics, not a test that distinguishes saturation from mathematical coupling. SV labels SSL labels Exp. EER^¯ EER maxEER EER r EER^¯ EER maxEER EER r E13 0.053 0.125 +0.070 0.254 0.399 +0.012 E14 0.039 0.351 +0.504 0.438 0.670 -0.706 E17 0.048 0.208 -0.220 0.392 0.467 -0.293 Across the five SSL labels in E14, mean predictions range from 0.397 to 0.457, observed means from 0.350 to 0.414, and mean contrasts from −0.073-0.073 to −0.028-0.028. This within-family variation is descriptive, rather than a control that separates competing explanations for negative contrasts. Demographic gaps under stress. E16 reports gender-stratified Common Voice conditions. We average the absolute per-model gap |EERF−EERM||EER_F-EER_M| over 12 model labels. The mean gap is 0.0086 in the clean condition. At 25, 15, and 5 dB, Gaussian-labelled noise gives 0.0191, 0.0129, and 0.0222; environmental-labelled noise gives 0.0182, 0.0230, and 0.0128; and crosstalk gives 0.0122, 0.0493, and 0.0278. All nine noisy means exceed the clean mean, but none of the three sequences increases monotonically as SNR falls (Figure 4). This does not imply an effect for every encoder or identify one population as consistently worse served. Appendix D.1 lists the subgroup summaries. Figure 4: E16 mean absolute gender EER gap by labelled SNR. Bars show SD across 12 model labels. Noise names follow the CSV; source-class provenance is unverified. The dashed line is the clean mean. Independently sampled trial lists and unequal clean/noisy budgets limit causal interpretation. The signed clean gap ranges in magnitude from 0.0004 to 0.0352 across models. Dividing noisy signed gaps by these small values produces ratios from −157.6-157.6 to +193.7+193.7 (Appendix D.2). Absolute gaps avoid that denominator, but still need sampling uncertainty: E16 uses 5,000 clean versus 555 noisy trials per subgroup, and absolute differences can increase with estimator variability. Three clean signed gaps are positive and nine negative, so there is no uniform direction across encoders. 6 Discussion An additive baseline makes the comparison of compound and single conditions explicit. The small signed means in the artefact motivate testing its predictive accuracy, but do not establish it. A useful validation would predefine an acceptable error, measure prediction errors on held-out factor settings, and retain their per-model distribution. A benchmark should report where the prediction fails, not only whether positive and negative deviations cancel. Metric interpretation is equally consequential. EER evaluates a condition’s own equal-error operating point; it does not directly measure risk at a fixed production threshold. Likewise, demographic EER gaps compare separately optimised operating points. Deployment-oriented analysis needs subgroup false-accept and false-reject rates at a common threshold selected on separate development data. These additions would test the practical meaning of the observed contrasts without treating benchmark averages as deployment guarantees. 7 Limitations and Threats to Validity Result provenance. The selected CSV includes run IDs for 2,508 records, revision identifiers for 2,130, checkpoint hashes for 1,878, and genuine-speaker counts for 924. Coverage is incomplete; trial lists, per-trial scores, hardware, and full software environments are not supplied. The legacy harness can substitute checkpoints under other labels, whereas the updated E13/E14 cohorts carry explicit hashes. Aggregate agreement alone cannot authenticate an encoder (Appendix F). Code–artefact discrepancies. The bundled legacy harness differs from producing revisions. The run audit found positive-gap execution, two-class E11 scoring, recorded E04 styles, real E05 codecs, and measured E20 reference terms. These findings supersede failures inferred solely from old code paths. Exact scored-pair manifests and full checkpoint binding are still unavailable; publication-year and age proxies, and the unrestricted nominal zero-gap reference, limit temporal interpretation (Appendix G). Uncertainty and comparability. Each retained record has one aggregate measurement and no linked score list. Repeated speakers and shared reference terms make the contrasts dependent. Between-model SDs do not quantify sampling uncertainty. Trial budgets vary, and corpus membership alone does not establish matched population composition. Scope. Per-trial noise, codec, synthesis, and conversion execution is not completely documented. E15/E18 mix source populations across corpora; E20 uses temporal and age proxies. Pre-training overlap, corpus versions, full subgroup sizes, and metadata provenance remain unaudited. These limits prevent broad claims about real-world compound robustness or demographic fairness. 8 Conclusion InterBias-SV organises compound-condition SV evaluation around measured reference terms and an explicit additive contrast. The supplied artefact permits recomputable summaries for three same-corpus experiment families, with small positive signed means for nominal speaker-trained labels and substantial setting-level variation. Those averages remain descriptive until trial matching and experimental provenance are established. Chance EER is not a hard ceiling, and a near-zero demographic reference makes gap ratios unstable. The framework and audit identify what must be measured and recorded before compound-condition results can support stronger conclusions. Ethics Statement The benchmark uses existing speech corpora and introduces no new participant recruitment or annotation. The supplementary package contains aggregate results and code, without corpus audio. Public availability does not itself establish consent for every downstream use: corpus versions, licence terms, and permitted uses must be checked with the original distributors. No claim of institutional ethics exemption follows from the absence of new data collection. Demographic categories are coarse, subgroup metadata are incomplete, and the retained gender and age comparisons do not establish fairness for specific populations. Speaker counts in some updated rows do not resolve label provenance or sampling uncertainty. The results should not be used to certify fairness or to justify deployment of biometric surveillance or access-control systems. Speech and linked demographic metadata can reveal identity even when speaker IDs are pseudonymous; any later release of manifests or audio-derived information needs a separate privacy and licence review. Spoofing and conversion experiments concern impersonation risks, but their execution must be verified before drawing vulnerability conclusions. No generated attack audio is included. Reproducibility Statement Section 3 defines scoring and contrasts; Section 4 and Appendix A.1 describe the available implementation. The supplementary CSV and two analysis scripts reconstruct aggregate tables and plots, with detailed listings in Appendices E and H. The protocol diagram is defined in the LaTeX source. This is summary-level reproducibility: Section 7 and Appendix G identify missing run provenance and code–artefact discrepancies that prevent full experimental reproduction. AI Use Statement Generative AI assisted implementation and analysis scripting, results-file formatting, interpretation of numerical summaries, literature retrieval, reference formatting, and manuscript drafting and editing. It also provided methodological criticism, suggested follow-up experiments, and helped revise mathematical statements about EER and interaction contrasts during the submission-readiness review. The supplied retained observations were not altered; this editorial review did not generate experimental observations or run new benchmark experiments. Checks in this revision covered aggregate recomputation, source-code consistency, selected primary references, and document compilation. These checks do not establish the provenance of the underlying experiments. The authors remain responsible for verifying the scientific content and for ensuring that the submission-form disclosure accurately reflects the full research workflow. References Ardila et al. (2020) R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. M. Tyers, and G. Weber Common voice: a massively-multilingual speech corpus. In Proc. Language Resources and Evaluation Conf. (LREC), p. 4218–4222. Cited by: §4. Baali et al. (2025) M. Baali, S. Bisht, F. Teixeira, K. Shapovalenko, R. Singh, and B. Raj SVeritas: benchmark for robust speaker verification under diverse conditions. In Findings of the Association for Computational Linguistics: EMNLP 2025, p. 9714–9731. Cited by: §1, §2. Baas et al. (2023) M. Baas, B. van Niekerk, and H. Kamper Voice conversion with just nearest neighbors. In Proc. Interspeech, p. 2053–2057. Cited by: Table 4, §4. Chen et al. (2022a) S. Chen, C. Wang, Z. Chen, Y. Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y. Qian, Y. Qian, J. Wu, M. Zeng, X. Yu, and F. Wei WavLM: large-scale self-supervised pre-training for full stack speech processing. IEEE J. Sel. Topics Signal Process. 16 (6), p. 1505–1518. Cited by: §2, §4. Chen et al. (2022b) S. Chen, Y. Wu, C. Wang, Z. Chen, Z. Chen, S. Liu, J. Wu, Y. Qian, F. Wei, J. Li, and X. Yu UniSpeech-SAT: universal speech representation learning with speaker aware pre-training. In Proc. IEEE ICASSP, p. 6152–6156. Cited by: §2, §4. Chung et al. (2018) J. S. Chung, A. Nagrani, and A. Zisserman VoxCeleb2: deep speaker recognition. In Proc. Interspeech, p. 1086–1090. Cited by: §1, §2. Desplanques et al. (2020) B. Desplanques, J. Thienpondt, and K. Demuynck ECAPA-TDNN: emphasized channel attention, propagation and aggregation in TDNN based speaker verification. In Proc. Interspeech, p. 3830–3834. Cited by: §4. Eskimez et al. (2024) S. E. Eskimez, X. Wang, M. Thakker, C. Li, C. Tsai, Z. Xiao, H. Yang, Z. Zhu, M. Tang, X. Tan, Y. Liu, S. Zhao, and N. Kanda E2 TTS: embarrassingly easy fully non-autoregressive zero-shot TTS. In Proc. IEEE Spoken Language Technology Workshop (SLT), Cited by: Table 4, §4. Fenu et al. (2021) G. Fenu, M. Marras, G. Medda, and G. Meloni Fair voice biometrics: impact of demographic imbalance on group fairness in speaker recognition. In Proc. Interspeech, p. 1892–1896. Cited by: §2. Hsu et al. (2021) W. Hsu, B. Bolte, Y. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed HuBERT: self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Trans. Audio, Speech, Language Process. 29, p. 3451–3460. Cited by: §2, §4. Huh et al. (2024) J. Huh, J. S. Chung, A. Nagrani, A. Brown, J. Jung, D. Garcia-Romero, and A. Zisserman The VoxCeleb speaker recognition challenge: a retrospective. IEEE/ACM Trans. Audio, Speech, Language Process. 32, p. 3850–3866. Cited by: §1, §2. Hutiri and Ding (2022) W. T. Hutiri and A. Y. Ding Bias in automated speaker recognition. In Proc. ACM Conf. Fairness, Accountability, and Transparency (FAccT), p. 230–247. Cited by: §2. Jung et al. (2022) J. Jung, Y. Kim, H. Heo, B. Lee, Y. Kwon, and J. S. Chung Pushing the limits of raw waveform speaker recognition. In Proc. Interspeech, p. 2228–2232. Cited by: §4. Ko et al. (2017) T. Ko, V. Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur A study on data augmentation of reverberant speech for robust speech recognition. In Proc. IEEE ICASSP, p. 5220–5224. Cited by: Table 4, §2, §4. Koenecke et al. (2020) A. Koenecke, A. Nam, E. Lake, J. Nudell, M. Quartey, Z. Mengesha, C. Toups, J. R. Rickford, D. Jurafsky, and S. Goel Racial disparities in automated speech recognition. Proc. Natl. Acad. Sci. U.S.A. 117 (14), p. 7684–7689. Cited by: §2. Koluguri et al. (2022) N. R. Koluguri, T. Park, and B. Ginsburg TitaNet: neural model for speaker representation with 1d depth-wise separable convolutions and global context. In Proc. IEEE ICASSP, p. 8102–8106. Cited by: §4. Li et al. (2023) J. Li, W. Tu, and L. Xiao FreeVC: towards high-quality text-free one-shot voice conversion. In Proc. IEEE ICASSP, Cited by: Table 4, §4. Livingstone and Russo (2018) S. R. Livingstone and F. A. Russo The Ryerson audio-visual database of emotional speech and song (RAVDESS): a dynamic, multimodal set of facial and vocal expressions in north american english. PLOS ONE 13 (5), p. e0196391. Cited by: §4. Nagrani et al. (2017) A. Nagrani, J. S. Chung, and A. Zisserman VoxCeleb: a large-scale speaker identification dataset. In Proc. Interspeech, p. 2616–2620. Cited by: §1, §2, §4. Peng et al. (2024) P. Peng, P. Huang, A. Mohamed, and D. Harwath VoiceCraft: zero-shot speech editing and text-to-speech in the wild. In Proc. Annu. Meeting Assoc. Comput. Linguistics (ACL), p. 12442–12462. Cited by: Table 4, §4. Peri et al. (2023) R. Peri, K. Somandepalli, and S. Narayanan A study of bias mitigation strategies for speaker recognition. Computer Speech & Language 79, p. 101481. Cited by: §2. Ravanelli et al. (2021) M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, J. Chou, S. Yeh, S. Fu, C. Liao, E. Rastorgueva, F. Grondin, W. Aris, H. Na, Y. Gao, R. De Mori, and Y. Bengio SpeechBrain: a general-purpose speech toolkit. arXiv:2106.04624. Cited by: §4. Richter et al. (2024) J. Richter, Y. Wu, S. Krenn, S. Welker, B. Lay, S. Watanabe, A. Richard, and T. Gerkmann EARS: an anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation. In Proc. Interspeech, p. 4873–4877. Cited by: §4. Rudzicz et al. (2012) F. Rudzicz, A. K. Namasivayam, and T. Wolff The TORGO database of acoustic and articulatory speech from speakers with dysarthria. Language Resources and Evaluation 46 (4), p. 523–541. Cited by: §4. Snyder et al. (2018) D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur X-vectors: robust DNN embeddings for speaker recognition. In Proc. IEEE ICASSP, p. 5329–5333. Cited by: §2, §4. Wang et al. (2023a) H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y. Deng, and Y. Qian Wespeaker: a research and production oriented speaker embedding learning toolkit. In Proc. IEEE ICASSP, Cited by: §4. Wang et al. (2023b) H. Wang, S. Zheng, Y. Chen, L. Cheng, and Q. Chen CAM++: a fast and efficient network for speaker verification using context-aware masking. In Proc. Interspeech, p. 5301–5305. Cited by: §4. Yakovlev et al. (2024) I. Yakovlev, R. Makarov, A. Balykin, P. Malov, A. Okhotnikov, and N. Torgashov Reshape dimensions network for speaker recognition. In Proc. Interspeech, p. 3235–3239. Cited by: §4. Yang et al. (2021) S. Yang, P. Chi, Y. Chuang, C. J. Lai, K. Lakhotia, Y. Y. Lin, A. T. Liu, J. Shi, X. Chang, G. Lin, T. Huang, W. Tseng, K. Lee, D. Liu, Z. Huang, S. Dong, S. Li, S. Watanabe, A. Mohamed, and H. Lee SUPERB: speech processing universal performance benchmark. In Proc. Interspeech, p. 1194–1198. Cited by: §2. Zhao et al. (2018) G. Zhao, S. Sonsaat, A. Silpachai, I. Lucic, E. Chukharev-Hudilainen, J. Levis, and R. Gutierrez-Osuna L2-ARCTIC: a non-native english speech corpus. In Proc. Interspeech, p. 2783–2787. Cited by: §4. Appendix A Benchmark Documentation A.1 Motivation, composition and intended use Purpose. InterBias-SV specifies additive comparisons between joint conditions and their measured reference terms. The results below reproduce the supplied aggregate artefact. Numerical availability is distinct from validated condition provenance and matched sampling (Section 3). Instances. A condition combines a source population, transformation, and trial-construction rule. The selected file contains 4,068 records with EER, minDCF, AUC, and trial counts across 17 experiments, totalling 11,975,026 trial evaluations. This sum includes reuse across conditions and models. All included records have finite scores; absent model–condition cells are not imputed. Degradation operators. Table 4 documents the available implementation, not proof that each operation succeeded in the runs that produced the CSV. Failure counts are supplied for some updated records, but execution provenance is incomplete. Table 4: Legacy harness operators (not the producing revisions for all retained rows). Proxy operators are marked p. Execution and checkpoint provenance are not verified by the result rows. Operator Settings Implementation Telephony codec G.711, GSM 06.10, AMR-NB, Opus ffmpeg encode–decode round trip; G.711 μ-law, GSM and AMR-NB at 8 kHz, Opus at 16 kHz / 12 kbit/s; failures return original audio Additive noise Gaussian, environmental, crosstalk; 5/15/25 dB SNR environmental and crosstalk labels share a RIRS_NOISES directory (Ko et al., 2017); missing files trigger Gaussian fallback Packet loss 0.5%, 2%, 5%, 10% 20 ms frames zeroed independently; no packet-loss concealment or burst-loss model Speaking rate 0.75×0.75×, 1.0×1.0×, 1.25×1.25× time-scale perturbation Wireless codecp 32–990 kbit/s every condition encoded with AAC at a varying bitrate under labels nominally denoting SBC, aptX and LDAC; measures perceptual-codec bitrate sensitivity, not codec identity Voice modep whisper, shout signal-level DSP proxy, not recorded phonation-mode speech; EARS contains genuine whispered and shouted recordings, which this operator does not use Synthesis VoiceCraft (Peng et al., 2024), E2-TTS (Eskimez et al., 2024) named wrappers; failures can return original audio Voice conversion kNN-VC (Baas et al., 2023), FreeVC (Li et al., 2023) named wrappers; failures can return original audio Audio front end 16 kHz; 1–10 s clips <1<1 s discarded, clips >10>10 s truncated Trials ≤104≤ 10^4 per condition balanced sampling requested; module-level seed 42; condition-specific resampling and exclusions; specialised samplers need separate checks Scoring cosine on L2L_2-normalised embeddings minDCFminDCF at Ptarget=0.01P_target=0.01, Cmiss=Cfa=1C_miss=C_fa=1; no score normalisation, calibration or adaptation Splits. There is no benchmark-specific training split or fine-tuning. Pre-trained encoders may nevertheless have seen evaluation speakers. Corpus versions, sampling manifests, and pre-training overlap need to be established before any condition can be described as held out. Intended use. The artefact supports descriptive condition summaries and recomputation of same-corpus contrasts. Controlled interaction, demographic, spoofing, and temporal-drift claims require the provenance and experimental checks described in Appendix G. A.2 Full experiment inventory with aggregate scores Table 5 is the per-experiment inventory: corpus, encoder and condition counts, scored records, trial counts, and the descriptive use of each experiment under Section 3. Table 6 adds the aggregate EER, minDCF and AUC per experiment. Those aggregates pool heterogeneous conditions and encoders: they describe the difficulty of a condition set and are not a controlled ranking of stressors. Table 5: Experiment inventory. “Scored” counts model–condition records with a computed EER. “Use” is what the experiment is admitted for: † cross-corpus populations: descriptive only; ‡ temporal/age proxy interpretation; § @sectionsign crosses a population attribute with a stressor, analysed as a subgroup gap; p legacy proxy implementation only. E05, E10, and E11 use operator labels rather than corpus identifiers in their dataset fields. All 12 retained encoders have complete coverage of each observed condition grid. T ID Factor(s) Corpus Enc. Cond. Scored Trials Use 1 E01 Emotional speech RAVDESS 12 15 180 330,000 marginal 1 E02 Dysarthric speech TORGO 12 4 48 153,706 marginal 1 E03 Speaking rate CommonVoice, EARS 12 18 216 2,159,910 marginal 1 E04 Recorded voice style EARS 12 9 108 117,036 marginal 1 E05 Codec round trips not recorded 12 15 180 1,799,550 marginal 1 E06 Packet loss CommonVoice, EARS 12 18 216 2,159,901 marginal 1 E07 Temporal drift VoxCeleb1 12 4 48 480,000 marginal 1 E08 L2 accent CommonVoice, L2-ARCTIC 12 18 216 378,840 marginal 1 E10 TTS spoofing not recorded 12 7 84 815,448 marginal 1 E11 Voice conversion not recorded 12 3 36 180,000 marginal 2 E13 Age × codec CommonVoice 12 20 240 1,199,940 contrast (audit) 2 E14 Emotion × noise RAVDESS 12 80 960 276,480 contrast (audit) 2 E15 Pathology × codec† EARS, TORGO 12 20 240 468,067 descriptive† 2 E16 Gender × noise CommonVoice 12 30 360 419,868 subgroup gap§ 2 E17 Drift × codec VoxCeleb1 12 20 240 479,760 contrast (audit) 2 E18 Accent × codec† CommonVoice, L2-ARCTIC 12 10 120 239,648 descriptive† 3 E20 Age × codec × drift‡ VoxCeleb1 12 48 576 316,872 descriptive‡ Table 6: Experiment inventory with aggregate scores. EER is the unweighted mean over scored records with its standard deviation across those records. † cross-corpus populations, reported descriptively; ‡ temporal/age proxy interpretation. Tier ID Factor(s) Enc. Cond. Scored Trials EER minDCF AUC 1 E01 Emotional speech 12 15 180 330,000 0.172 ± 0.137 0.617 0.880 1 E02 Dysarthric speech 12 4 48 153,706 0.318 ± 0.109 0.964 0.737 1 E03 Speaking rate 12 18 216 2,159,910 0.351 ± 0.132 0.887 0.688 1 E04 Recorded voice style 12 9 108 117,036 0.182 ± 0.183 0.531 0.850 1 E05 Codec round trips 12 15 180 1,799,550 0.252 ± 0.161 0.686 0.791 1 E06 Packet loss 12 18 216 2,159,901 0.188 ± 0.162 0.582 0.853 1 E07 Temporal drift 12 4 48 480,000 0.175 ± 0.183 0.524 0.854 1 E08 L2 accent 12 18 216 378,840 0.136 ± 0.169 0.413 0.891 1 E10 TTS spoofing 12 7 84 815,448 0.407 ± 0.147 0.876 0.620 1 E11 Voice conversion 12 3 36 180,000 0.274 ± 0.128 0.737 0.774 2 E13 Age × codec 12 20 240 1,199,940 0.130 ± 0.110 0.505 0.913 2 E14 Emotion × noise 12 80 960 276,480 0.178 ± 0.195 0.439 0.846 2 E15 Pathology × codec† 12 20 240 468,067 0.311 ± 0.119 0.883 0.739 2 E16 Gender × noise 12 30 360 419,868 0.182 ± 0.156 0.540 0.863 2 E17 Drift × codec 12 20 240 479,760 0.188 ± 0.177 0.560 0.847 2 E18 Accent × codec† 12 10 120 239,648 0.145 ± 0.144 0.529 0.893 3 E20 Age × codec × drift‡ 12 48 576 316,872 0.186 ± 0.176 0.524 0.850 Historical descriptive experiments (E15 and E18). E15 contributes 240 scored rows (12 encoders, 20 conditions): TORGO dysarthric clean/codec results and EARS healthy-reference results. E18 contributes 120 rows (12 encoders, 10 conditions): L2-ARCTIC and Common Voice English clean/codec results. All condition grids are complete. Compare clean and codec conditions within each corpus; differences between corpora cannot isolate pathology or accent. E18’s literal native condition labels are retained for lookup, but English locale does not establish native-speaker status. These historical rows enter descriptive summaries only, not the E13/E14/E17 interaction table. Appendix B Encoder Results Table 7 gives macro-average results: conditions are averaged within each experiment, then experiments are weighted equally. The last column is the number of experiments contributing: seventeen for every retained encoder. Condition coverage is complete, but trial matching and provenance limitations keep the ordering descriptive. Table 7: Macro-average encoder results. “Exp.” is the number of experiments contributing to that encoder’s macro average. Encoder Family EER minDCF AUC Exp. CAM++ SV 0.098 0.347 0.933 17 ECAPA-TDNN SV 0.101 0.360 0.932 17 ECAPA-TDNN-Large SV 0.102 0.357 0.931 17 Titanet SV 0.104 0.341 0.928 17 RawNet3 SV 0.113 0.409 0.917 17 RedimNet SV 0.115 0.381 0.919 17 X-Vector SV 0.181 0.637 0.877 17 WavLM-Large SSL 0.331 0.949 0.720 17 WavLM-Base SSL 0.350 0.927 0.695 17 WavLM-Base+ SSL 0.371 0.961 0.673 17 UniSpeech-SAT SSL 0.380 0.967 0.661 17 HuBERT-Large SSL 0.415 0.986 0.617 17 Appendix C Single-Factor Results: E01 Emotional Speech Table 8: E01, RAVDESS. Mean EER over available encoders, with the standard deviation across encoders; the last column is the ratio to the neutral/neutral reference. “Matched” conditions use the same emotion for enrolment and test. The SD is between-encoder variability, not a confidence interval. All conditions cover the same 12 model labels. Ratios divide the condition mean by the neutral-reference mean; they do not quantify sampling uncertainty. Condition Type EER SD minDCF AUC Ratio neutral/neutral matched 0.083 0.114 0.424 0.951 1.00× calm/calm matched 0.106 0.119 0.378 0.936 1.28× surprise/surprise matched 0.116 0.117 0.501 0.927 1.40× disgust/disgust matched 0.131 0.127 0.513 0.918 1.59× neutral→ mismatched 0.136 0.137 0.527 0.905 1.65× happy/happy matched 0.151 0.128 0.580 0.899 1.84× neutral→ mismatched 0.170 0.148 0.709 0.879 2.06× sad/sad matched 0.177 0.118 0.603 0.881 2.14× fearful/fearful matched 0.191 0.120 0.589 0.869 2.32× angry/angry matched 0.199 0.119 0.671 0.870 2.41× neutral→ mismatched 0.200 0.139 0.704 0.856 2.42× neutral→ mismatched 0.210 0.154 0.768 0.842 2.54× neutral→ mismatched 0.210 0.160 0.711 0.843 2.55× neutral→ mismatched 0.238 0.140 0.777 0.819 2.89× neutral→ mismatched 0.256 0.155 0.798 0.799 3.11× Appendix D Demographic Results D.1 Absolute gender gap by condition Table 9: E16 gender gap on Common Voice. F and M are the recorded subgroup labels; columns give mean EERs over 12 model labels; ||gap|| is the mean absolute per-encoder difference, and max||gap|| its largest value over encoders. Condition EER (F) EER (M) ||gap|| max||gap|| Enc. clean (reference) 0.110 0.117 0.0086 0.0352 12 Gaussian, 25 dB 0.139 0.141 0.0191 0.0595 12 Gaussian, 15 dB 0.139 0.146 0.0129 0.0432 12 Gaussian, 5 dB 0.159 0.180 0.0222 0.0450 12 environmental, 25 dB 0.131 0.147 0.0182 0.0811 12 environmental, 15 dB 0.171 0.187 0.0230 0.0523 12 environmental, 5 dB 0.201 0.200 0.0128 0.0631 12 crosstalk, 25 dB 0.167 0.162 0.0122 0.0324 12 crosstalk, 15 dB 0.270 0.236 0.0493 0.0829 12 crosstalk, 5 dB 0.347 0.344 0.0278 0.0883 12 D.2 Instability of ratio-form bias amplification Table 10: Instability of ratio-form bias amplification on E16. The clean gender gap is the denominator of the ratio; “mean ||gap||” is the mean absolute gap over the nine noise conditions. The ratio range is reported to show its sensitivity to a near-zero denominator, not as a reliable amplification estimate. Encoder Clean gap Mean ||gap|| Ratio range CAM++ -0.0022 0.0208 -37.7 to +8.2 ECAPA-TDNN -0.0076 0.0158 -5.2 to +3.8 ECAPA-TDNN-Large -0.0106 0.0194 -6.3 to +1.9 HuBERT-Large +0.0036 0.0198 -17.5 to +3.0 RawNet3 -0.0048 0.0164 -6.4 to +6.8 RedimNet +0.0012 0.0188 -25.5 to +43.5 Titanet -0.0020 0.0146 -13.5 to +15.3 UniSpeech-SAT -0.0216 0.0178 -1.5 to +1.4 WavLM-Base -0.0126 0.0499 -7.0 to +6.4 WavLM-Base+ -0.0352 0.0252 -1.7 to +1.5 WavLM-Large -0.0004 0.0218 -157.6 to +193.7 X-Vector +0.0014 0.0226 -28.3 to +19.3 Appendix E Interaction Analysis in Detail E.1 Per-encoder interaction effects Table 11 decomposes the contrasts by model label. In E14, all five SSL means are negative, ranging from −0.073-0.073 to −0.028-0.028 EER. Six of seven speaker-trained means are positive; RawNet3 is negative (−0.003-0.003). The table includes all finite four-term contrasts, while NLI means use only positive additive predictions. No family-level chance screen is applied; sampling and provenance limitations remain. Table 11: Interaction effect by encoder and experiment, over the settings for which all four factorial terms are available. n is the number of settings. Exp. Encoder Fam. IE¯ IE SD min max NLI¯ NLI n E13 CAM++ SV +0.0028 0.0103 -0.0127 +0.0248 1.226 12 E13 ECAPA-TDNN SV +0.0019 0.0141 -0.0154 +0.0218 1.020 12 E13 ECAPA-TDNN-Large SV -0.0050 0.0158 -0.0467 +0.0112 0.940 12 E13 HuBERT-Large SSL +0.0003 0.0334 -0.0668 +0.0520 1.007 12 E13 RawNet3 SV -0.0046 0.0178 -0.0491 +0.0240 0.949 12 E13 RedimNet SV +0.0087 0.0116 -0.0119 +0.0297 1.356 12 E13 Titanet SV +0.0005 0.0129 -0.0151 +0.0263 1.016 12 E13 UniSpeech-SAT SSL -0.0003 0.0154 -0.0430 +0.0126 1.000 12 E13 WavLM-Base SSL -0.0136 0.0210 -0.0730 +0.0020 0.931 12 E13 WavLM-Base+ SSL -0.0100 0.0131 -0.0480 +0.0033 0.962 12 E13 WavLM-Large SSL -0.0077 0.0149 -0.0453 +0.0060 0.965 12 E13 X-Vector SV +0.0140 0.0407 -0.0231 +0.1011 1.118 12 E14 CAM++ SV +0.0084 0.0225 +0.0000 +0.1076 1.275 63 E14 ECAPA-TDNN SV +0.0134 0.0252 -0.0243 +0.1250 1.355 63 E14 ECAPA-TDNN-Large SV +0.0119 0.0358 -0.0069 +0.1389 1.284 63 E14 HuBERT-Large SSL -0.0364 0.0445 -0.1250 +0.0799 0.927 63 E14 RawNet3 SV -0.0031 0.0331 -0.0521 +0.1076 0.564 63 E14 RedimNet SV +0.0161 0.0320 -0.0069 +0.1146 6.753 63 E14 Titanet SV +0.0138 0.0325 +0.0000 +0.1285 1.484 63 E14 UniSpeech-SAT SSL -0.0728 0.0531 -0.1701 +0.0347 0.860 63 E14 WavLM-Base SSL -0.0469 0.0551 -0.1562 +0.0833 0.954 63 E14 WavLM-Base+ SSL -0.0279 0.0502 -0.1215 +0.0972 0.962 63 E14 WavLM-Large SSL -0.0484 0.0572 -0.1285 +0.0799 0.920 63 E14 X-Vector SV +0.0014 0.0230 -0.0590 +0.0451 0.911 63 E17 CAM++ SV +0.0007 0.0047 -0.0071 +0.0096 1.034 12 E17 ECAPA-TDNN SV +0.0044 0.0059 -0.0034 +0.0133 1.204 12 E17 ECAPA-TDNN-Large SV +0.0071 0.0093 -0.0103 +0.0194 1.183 12 E17 HuBERT-Large SSL -0.0030 0.0224 -0.0484 +0.0284 0.995 12 E17 RawNet3 SV +0.0063 0.0093 -0.0064 +0.0309 1.296 12 E17 RedimNet SV +0.0020 0.0106 -0.0129 +0.0244 1.022 12 E17 Titanet SV +0.0019 0.0048 -0.0055 +0.0087 1.265 12 E17 UniSpeech-SAT SSL +0.0186 0.0156 -0.0174 +0.0362 1.049 12 E17 WavLM-Base SSL +0.0074 0.0200 -0.0217 +0.0299 1.019 12 E17 WavLM-Base+ SSL -0.0012 0.0201 -0.0252 +0.0365 0.997 12 E17 WavLM-Large SSL +0.0028 0.0155 -0.0276 +0.0286 1.009 12 E17 X-Vector SV -0.0055 0.0138 -0.0255 +0.0175 0.957 12 E.2 Setting-level interaction listing Table 12 gives the pooled observed joint EER and additive prediction, followed by separate family-mean contrasts for every setting with four available terms. The dispersion visible here is what the pooled means of Section 5 average over. Table 12: Per-setting interaction summary. “Obs.” and “Add.” pool all available encoder labels; the final two columns give family-specific mean contrasts. Exp. Setting Obs. (all) Add. (all) IEIE (SV) IEIE (SSL) E13 middle amr 0.153 0.149 0.0025 0.0053 E13 middle g711 0.148 0.143 0.0016 0.0089 E13 middle gsm 0.158 0.153 0.0047 0.0054 E13 middle opus 0.134 0.137 -0.0019 -0.0055 E13 older amr 0.141 0.146 0.0117 -0.0265 E13 older g711 0.135 0.140 0.0016 -0.0132 E13 older gsm 0.143 0.149 0.0220 -0.0444 E13 older opus 0.132 0.134 -0.0048 0.0009 E13 young amr 0.123 0.126 -0.0063 -0.0004 E13 young g711 0.121 0.120 0.0003 0.0016 E13 young gsm 0.127 0.130 0.0006 -0.0065 E13 young opus 0.114 0.115 -0.0005 -0.0006 E14 ravdess angry crosstalk 15dB 0.267 0.269 0.0253 -0.0396 E14 ravdess angry crosstalk 25dB 0.214 0.229 0.0045 -0.0410 E14 ravdess angry crosstalk 5dB 0.380 0.367 0.0615 -0.0563 E14 ravdess angry environmental 15dB 0.182 0.202 0.0005 -0.0500 E14 ravdess angry environmental 25dB 0.160 0.194 -0.0015 -0.0792 E14 ravdess angry environmental 5dB 0.196 0.219 0.0015 -0.0590 E14 ravdess angry gaussian 15dB 0.112 0.102 -0.0050 0.0312 E14 ravdess angry gaussian 25dB 0.094 0.096 -0.0055 0.0028 E14 ravdess angry gaussian 5dB 0.136 0.114 0.0005 0.0528 E14 ravdess calm crosstalk 15dB 0.239 0.262 -0.0050 -0.0486 E14 ravdess calm crosstalk 25dB 0.194 0.222 -0.0010 -0.0674 E14 ravdess calm crosstalk 5dB 0.371 0.361 0.0437 -0.0375 E14 ravdess calm environmental 15dB 0.175 0.196 -0.0010 -0.0479 E14 ravdess calm environmental 25dB 0.156 0.187 -0.0045 -0.0687 E14 ravdess calm environmental 5dB 0.196 0.213 0.0020 -0.0424 E14 ravdess calm gaussian 15dB 0.097 0.096 -0.0040 0.0076 E14 ravdess calm gaussian 25dB 0.067 0.090 -0.0040 -0.0479 E14 ravdess calm gaussian 5dB 0.098 0.108 -0.0025 -0.0187 E14 ravdess disgust crosstalk 15dB 0.255 0.285 0.0263 -0.1090 E14 ravdess disgust crosstalk 25dB 0.209 0.245 0.0040 -0.0917 E14 ravdess disgust crosstalk 5dB 0.387 0.383 0.0685 -0.0868 E14 ravdess disgust environmental 15dB 0.178 0.218 0.0005 -0.0979 E14 ravdess disgust environmental 25dB 0.165 0.209 -0.0035 -0.1014 E14 ravdess disgust environmental 5dB 0.204 0.235 0.0040 -0.0812 E14 ravdess disgust gaussian 15dB 0.122 0.118 -0.0040 0.0153 E14 ravdess disgust gaussian 25dB 0.124 0.112 -0.0060 0.0361 E14 ravdess disgust gaussian 5dB 0.120 0.130 -0.0025 -0.0215 E14 ravdess fearful crosstalk 15dB 0.256 0.263 0.0322 -0.0632 E14 ravdess fearful crosstalk 25dB 0.213 0.223 0.0129 -0.0431 E14 ravdess fearful crosstalk 5dB 0.401 0.362 0.0908 -0.0340 E14 ravdess fearful environmental 15dB 0.172 0.197 0.0000 -0.0590 E14 ravdess fearful environmental 25dB 0.149 0.188 -0.0010 -0.0917 E14 ravdess fearful environmental 5dB 0.194 0.214 -0.0000 -0.0479 E14 ravdess fearful gaussian 15dB 0.108 0.097 -0.0015 0.0299 E14 ravdess fearful gaussian 25dB 0.095 0.091 -0.0015 0.0132 E14 ravdess fearful gaussian 5dB 0.119 0.109 0.0030 0.0215 E14 ravdess happy crosstalk 15dB 0.247 0.280 0.0079 -0.0889 E14 ravdess happy crosstalk 25dB 0.216 0.240 0.0055 -0.0653 E14 ravdess happy crosstalk 5dB 0.369 0.378 0.0382 -0.0757 E14 ravdess happy environmental 15dB 0.177 0.213 -0.0010 -0.0868 E14 ravdess happy environmental 25dB 0.159 0.205 -0.0020 -0.1069 E14 ravdess happy environmental 5dB 0.201 0.230 -0.0015 -0.0674 E14 ravdess happy gaussian 15dB 0.105 0.113 -0.0045 -0.0132 E14 ravdess happy gaussian 25dB 0.096 0.107 -0.0045 -0.0201 E14 ravdess happy gaussian 5dB 0.122 0.125 -0.0020 -0.0042 E14 ravdess sad crosstalk 15dB 0.254 0.281 0.0099 -0.0785 E14 ravdess sad crosstalk 25dB 0.203 0.241 0.0010 -0.0931 E14 ravdess sad crosstalk 5dB 0.381 0.380 0.0506 -0.0674 E14 ravdess sad environmental 15dB 0.175 0.214 -0.0055 -0.0861 E14 ravdess sad environmental 25dB 0.161 0.206 -0.0050 -0.1007 E14 ravdess sad environmental 5dB 0.194 0.231 -0.0045 -0.0847 E14 ravdess sad gaussian 15dB 0.115 0.114 -0.0079 0.0139 E14 ravdess sad gaussian 25dB 0.107 0.108 -0.0079 0.0083 E14 ravdess sad gaussian 5dB 0.114 0.126 -0.0050 -0.0215 E14 ravdess surprise crosstalk 15dB 0.276 0.284 0.0536 -0.0951 E14 ravdess surprise crosstalk 25dB 0.229 0.244 0.0228 -0.0674 E14 ravdess surprise crosstalk 5dB 0.397 0.383 0.0873 -0.0889 E14 ravdess surprise environmental 15dB 0.188 0.218 0.0005 -0.0708 E14 ravdess surprise environmental 25dB 0.172 0.209 -0.0020 -0.0868 E14 ravdess surprise environmental 5dB 0.205 0.235 0.0055 -0.0785 E14 ravdess surprise gaussian 15dB 0.120 0.117 -0.0020 0.0090 E14 ravdess surprise gaussian 25dB 0.115 0.111 -0.0050 0.0146 E14 ravdess surprise gaussian 5dB 0.132 0.129 0.0064 -0.0035 E17 1yr amr 0.202 0.193 -0.0007 0.0235 E17 1yr g711 0.196 0.187 -0.0030 0.0254 E17 1yr gsm 0.206 0.203 -0.0031 0.0120 E17 1yr opus 0.183 0.177 -0.0053 0.0220 E17 3yr amr 0.198 0.197 0.0098 -0.0105 E17 3yr g711 0.192 0.191 0.0052 -0.0069 E17 3yr gsm 0.209 0.207 0.0155 -0.0160 E17 3yr opus 0.178 0.181 0.0006 -0.0089 E17 5yr amr 0.199 0.193 0.0048 0.0075 E17 5yr g711 0.190 0.187 -0.0007 0.0075 E17 5yr gsm 0.209 0.203 0.0089 0.0021 E17 5yr opus 0.176 0.177 -0.0029 0.0013 Appendix F Encoder-Agreement Audit Several model wrappers can replace a requested encoder with a SpeechBrain ECAPA-TDNN checkpoint, without recording the resolved identity in the CSV. Table 13 compares aggregate rows. No pair agrees exactly on more than 20.9% of the compared conditions, but this cannot exclude fallback: models can receive different sampled trials and augmentation realisations. Agreement rates include exactly zero errors in the updated matched conditions and cannot authenticate a checkpoint. The provided hashes offer additional provenance for the subset of records that contains them. Table 13: The twelve most similar encoder pairs of 66, by exact agreement over co-observed conditions. “Cond.” is the number of co-observed conditions, “Agree” the percentage on which the two EERs are identical, ρ the correlation of their EERs, and MAD the mean absolute EER difference. Encoder A Encoder B Cond. Agree (%) ρ MAD CAM++ ECAPA-TDNN-Large 339 20.9 0.993 0.0075 CAM++ Titanet 339 20.6 0.984 0.0126 ECAPA-TDNN-Large Titanet 339 19.5 0.982 0.0139 ECAPA-TDNN-Large RedimNet 339 18.0 0.919 0.0184 ECAPA-TDNN Titanet 339 16.5 0.846 0.0270 CAM++ RedimNet 339 16.2 0.921 0.0187 RedimNet Titanet 339 16.2 0.926 0.0192 CAM++ RawNet3 339 15.0 0.969 0.0191 CAM++ ECAPA-TDNN 339 14.7 0.829 0.0296 RawNet3 Titanet 339 14.7 0.973 0.0171 ECAPA-TDNN RedimNet 339 14.2 0.812 0.0326 ECAPA-TDNN ECAPA-TDNN-Large 339 13.9 0.816 0.0309 Appendix G Extended Limitations Uncertainty. A single run does not in principle prevent uncertainty estimation from retained scores. Here, only aggregate measurements are supplied. Recover per-trial scores and identities, then use a sampling scheme that respects repeated speakers and shared references across the four terms. Independent speaker/trial and augmentation seeds would separately test construction sensitivity. Treating all model–setting combinations as independent replicates would understate dependence. An equivalence claim additionally needs a predefined practical margin; a small signed mean is insufficient. Checkpoint and run provenance. Preserve the resolved model identifier, checkpoint hash, software environment, and successful operator configuration for each run. Disable silent substitution in future measurements. The updated summary includes run identifiers, revision fields, and checkpoint hashes for subsets of records, but no score manifests linking all observations to the bundled legacy implementation. The agreement table cannot fill this gap. Temporal labels and producing revisions. The run audit found per-video years, positive-gap execution and measured E20 reference terms. The old bundled loader is not the producing revision. Publication dates and age estimates remain proxies; the nominal zero-gap reference is unrestricted, not a verified same-session or same-year baseline. Complete pair manifests are needed to verify sampling and overlap. Noise and attack operators. The audit found two-class E11 scoring and stored synthesis/conversion outputs. This resolves contradictions inferred from the legacy one-class branch, but does not supply exact scored-pair lists or establish real-time latency. Noise-source corrections and failure counts cover some runs; complete operator and checkpoint binding remains unavailable. Demographic and emotion labels. E13 compares an age stratum with an overlapping all-age reference, not two disjoint age populations. Common Voice age bins are mapped to representative integers rather than measured ages. Report speaker counts, missing labels, label sources, and mapping rules for all strata. The retained E01 and replacement E14 rows use RAVDESS. The paired and matched cohort labels do not replace trial manifests needed to verify population and content matching. Contamination and coverage. Check pre-training/evaluation speaker overlap for each resolved checkpoint, especially on VoxCeleb1. All twelve retained labels cover all seventeen experiments and every observed condition. Comparable speaker composition and successful trials still need verification; complete coverage does not establish matched sampling. Excluded experiments and cohorts. E21 is omitted: only one model has an updated cohort, with the other twelve represented by legacy rows. The single-model legacy E13 cohort is omitted in favour of the 12-model paired replacement. E09, E12, and E19 have no valid replacement results. ERes2Net is excluded because it lacks four retained experiments. E15 and E18 are restored only as historical descriptive results. E20 remains descriptive and is not used as a three-way interaction estimate. Scale and release metadata. The trial total counts scoring operations, not independent data. The release lacks complete corpus versions, speaker-count coverage, audio hours, run hardware, exact library versions, and execution costs. Some records include checkpoint hashes and speaker counts. E05, E10, and E11 lack usable corpus identifiers. An artefact licence, third-party attribution, acquisition instructions, and stable protocol manifests are needed for reuse. Appendix H Condition-Level Results for Every Experiment The tables below list every scored condition in the selected 17 experiments, sorted by mean EER. “SD” is the standard deviation across contributing model rows, not a confidence interval. There are no unscored rows in the selected file. Every observed condition has all twelve retained encoders; the encoder count in each row makes this coverage explicit. CV denotes Common Voice, and L2-ARCTIC/CV denotes a condition involving both corpora. The E13 and E14 identifiers correspond to the paired-age and matched-RAVDESS replacement exports. Their double-underscore condition strings are preserved for exact lookup. The run audit identifies recorded EARS styles for E04 and actual codec round trips for E05; the proxy paths in the bundled legacy harness do not describe those retained observations. H.1 E01 – Emotional speech Table 14: E01 – Emotional speech: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials ravdess/neutral_neutral ravdess 12 0.083 0.114 0.424 0.951 120,000 ravdess/calm_calm ravdess 12 0.106 0.119 0.378 0.936 15,000 ravdess/surprise_surprise ravdess 12 0.116 0.117 0.501 0.927 15,000 ravdess/disgust_disgust ravdess 12 0.131 0.127 0.513 0.918 15,000 ravdess/neutral_calm ravdess 12 0.136 0.137 0.527 0.905 15,000 ravdess/happy_happy ravdess 12 0.151 0.128 0.580 0.899 15,000 ravdess/neutral_surprise ravdess 12 0.170 0.148 0.709 0.879 15,000 ravdess/sad_sad ravdess 12 0.177 0.118 0.603 0.881 15,000 ravdess/fearful_fearful ravdess 12 0.191 0.120 0.589 0.869 15,000 ravdess/angry_angry ravdess 12 0.199 0.119 0.671 0.870 15,000 ravdess/neutral_sad ravdess 12 0.200 0.139 0.704 0.856 15,000 ravdess/neutral_happy ravdess 12 0.210 0.154 0.768 0.842 15,000 ravdess/neutral_disgust ravdess 12 0.210 0.160 0.711 0.843 15,000 ravdess/neutral_angry ravdess 12 0.238 0.140 0.777 0.819 15,000 ravdess/neutral_fearful ravdess 12 0.256 0.155 0.798 0.799 15,000 H.2 E02 – Dysarthric speech Table 15: E02 – Dysarthric speech: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials TORGO/dysarthria_moderate_matched torgo 12 0.302 0.117 0.941 0.755 37,404 TORGO/control_matched torgo 12 0.318 0.120 0.962 0.735 39,689 TORGO/dysarthria_mild_matched torgo 12 0.326 0.105 0.978 0.730 38,154 TORGO/dysarthria_severe_matched torgo 12 0.327 0.106 0.977 0.726 38,459 H.3 E03 – Speaking rate Table 16: E03 – Speaking rate: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials common_voice/clean CV 12 0.108 0.118 0.454 0.933 120,000 common_voice/slow_slow CV 12 0.167 0.077 0.767 0.899 120,000 common_voice/fast_fast CV 12 0.192 0.070 0.798 0.881 120,000 ears/clean ears 12 0.244 0.174 0.658 0.796 119,990 common_voice/fast_normal CV 12 0.336 0.091 0.927 0.718 120,000 common_voice/normal_fast CV 12 0.337 0.090 0.928 0.718 120,000 common_voice/normal_slow CV 12 0.350 0.093 0.931 0.699 120,000 common_voice/slow_normal CV 12 0.351 0.093 0.930 0.698 120,000 ears/slow_slow ears 12 0.360 0.044 0.938 0.691 119,990 common_voice/fast_slow CV 12 0.379 0.088 0.936 0.661 120,000 common_voice/slow_fast CV 12 0.381 0.086 0.936 0.661 120,000 ears/fast_fast ears 12 0.381 0.032 0.966 0.664 119,990 ears/fast_normal ears 12 0.450 0.067 0.966 0.566 119,990 ears/normal_fast ears 12 0.451 0.067 0.962 0.566 119,990 ears/slow_normal ears 12 0.452 0.059 0.957 0.565 119,990 ears/normal_slow ears 12 0.452 0.060 0.956 0.565 119,990 ears/slow_fast ears 12 0.461 0.038 0.974 0.555 119,990 ears/fast_slow ears 12 0.463 0.041 0.977 0.552 119,990 H.4 E04 – Recorded voice style Table 17: E04 – Recorded voice style: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials shout_shout ears 12 0.148 0.183 0.428 0.870 13,332 modal_modal ears 12 0.148 0.182 0.424 0.875 13,332 whisper_whisper ears 12 0.157 0.165 0.477 0.880 13,332 modal_shout ears 12 0.162 0.189 0.441 0.864 12,840 shout_modal ears 12 0.171 0.201 0.449 0.853 12,840 whisper_modal ears 12 0.209 0.194 0.626 0.827 12,840 modal_whisper ears 12 0.209 0.193 0.657 0.831 12,840 whisper_shout ears 12 0.212 0.191 0.632 0.826 12,840 shout_whisper ears 12 0.217 0.194 0.645 0.824 12,840 H.5 E05 – Codec round trips Table 18: E05 – Codec round trips: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials aac_high_matched – 12 0.246 0.172 0.669 0.797 119,970 clean – 12 0.246 0.172 0.660 0.796 119,970 aac_medium_matched – 12 0.246 0.171 0.671 0.797 119,970 aptx_high_matched – 12 0.246 0.172 0.662 0.796 119,970 sbc_high_matched – 12 0.246 0.172 0.669 0.796 119,970 aptx_vs_aac_high – 12 0.246 0.171 0.671 0.796 119,970 aac_vs_aptx_medium – 12 0.246 0.173 0.666 0.796 119,970 aptx_low_matched – 12 0.248 0.170 0.661 0.795 119,970 sbc_vs_aac_medium – 12 0.250 0.168 0.686 0.793 119,970 sbc_vs_aptx_medium – 12 0.250 0.168 0.681 0.793 119,970 aptx_vs_sbc_high – 12 0.251 0.168 0.684 0.793 119,970 aac_vs_sbc_medium – 12 0.251 0.168 0.682 0.793 119,970 sbc_medium_matched – 12 0.252 0.166 0.687 0.791 119,970 aac_low_matched – 12 0.254 0.165 0.725 0.792 119,970 sbc_low_matched – 12 0.308 0.121 0.819 0.743 119,970 H.6 E06 – Packet loss Table 19: E06 – Packet loss: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials common_voice/clean CV 12 0.108 0.117 0.463 0.933 120,000 common_voice/clean_x_0.5pct CV 12 0.109 0.118 0.462 0.931 120,000 common_voice/0.5pct_0.5pct CV 12 0.110 0.118 0.463 0.930 120,000 common_voice/clean_x_2.0pct CV 12 0.112 0.121 0.473 0.929 120,000 common_voice/2.0pct_2.0pct CV 12 0.112 0.121 0.474 0.928 120,000 common_voice/5.0pct_5.0pct CV 12 0.116 0.123 0.482 0.925 120,000 common_voice/clean_x_5.0pct CV 12 0.116 0.124 0.478 0.924 120,000 common_voice/10.0pct_10.0pct CV 12 0.120 0.125 0.495 0.922 120,000 common_voice/clean_x_10.0pct CV 12 0.124 0.131 0.500 0.917 120,000 ears/0.5pct_0.5pct ears 12 0.245 0.172 0.662 0.798 119,989 ears/2.0pct_2.0pct ears 12 0.245 0.172 0.666 0.798 119,989 ears/5.0pct_5.0pct ears 12 0.246 0.172 0.667 0.798 119,989 ears/10.0pct_10.0pct ears 12 0.248 0.170 0.677 0.796 119,989 ears/clean ears 12 0.257 0.165 0.667 0.787 119,989 ears/clean_x_0.5pct ears 12 0.278 0.178 0.704 0.762 119,989 ears/clean_x_2.0pct ears 12 0.278 0.177 0.711 0.762 119,989 ears/clean_x_5.0pct ears 12 0.278 0.178 0.709 0.762 119,989 ears/clean_x_10.0pct ears 12 0.280 0.177 0.719 0.760 119,989 H.7 E07 – Temporal drift Table 20: E07 – Temporal drift: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials voxceleb1/gap_0yr voxceleb1 12 0.168 0.184 0.501 0.862 120,000 voxceleb1/gap_5yr voxceleb1 12 0.176 0.188 0.542 0.855 120,000 voxceleb1/gap_1yr voxceleb1 12 0.177 0.190 0.522 0.852 120,000 voxceleb1/gap_3yr voxceleb1 12 0.180 0.193 0.533 0.848 120,000 H.8 E08 – L2 accent Table 21: E08 – L2 accent: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials l2_arctic/level_3_vs_native L2-ARCTIC/CV 12 0.077 0.103 0.372 0.956 19,944 l2_arctic/level_1_vs_native L2-ARCTIC/CV 12 0.079 0.105 0.386 0.954 19,908 l2_arctic/level_4_vs_native L2-ARCTIC/CV 12 0.082 0.117 0.350 0.948 19,968 l2_arctic/l2_all_vs_native L2-ARCTIC/CV 12 0.084 0.116 0.371 0.951 19,932 l2_arctic/level_2_vs_native L2-ARCTIC/CV 12 0.088 0.116 0.376 0.944 19,932 l2_arctic/level_5_vs_native L2-ARCTIC/CV 12 0.091 0.120 0.388 0.944 19,932 cv_en/native_matched CV 12 0.110 0.118 0.411 0.935 19,992 l2_arctic/level_3_matched L2-ARCTIC 12 0.142 0.171 0.430 0.888 23,892 l2_arctic/level_4_matched L2-ARCTIC 12 0.150 0.184 0.416 0.872 23,964 l2_arctic/arabic_matched L2-ARCTIC 12 0.159 0.191 0.418 0.863 19,968 l2_arctic/level_2_matched L2-ARCTIC 12 0.165 0.196 0.440 0.861 23,916 l2_arctic/level_1_matched L2-ARCTIC 12 0.168 0.205 0.433 0.851 23,904 l2_arctic/spanish_matched L2-ARCTIC 12 0.168 0.201 0.430 0.855 19,932 l2_arctic/korean_matched L2-ARCTIC 12 0.169 0.209 0.421 0.845 19,980 l2_arctic/level_5_matched L2-ARCTIC 12 0.170 0.198 0.462 0.852 23,904 l2_arctic/vietnamese_matched L2-ARCTIC 12 0.176 0.203 0.463 0.848 19,920 l2_arctic/mandarin_matched L2-ARCTIC 12 0.181 0.216 0.428 0.840 19,944 l2_arctic/hindi_matched L2-ARCTIC 12 0.182 0.209 0.445 0.834 19,908 H.9 E10 – TTS spoofing Table 22: E10 – TTS spoofing: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials real_real_baseline – 12 0.239 0.170 0.649 0.804 120,000 voicecraft/cross_utt_spoof – 12 0.311 0.094 0.781 0.750 113,664 e2_tts/cross_utt_spoof – 12 0.368 0.050 0.913 0.691 118,104 voicecraft/synth_vs_real – 12 0.433 0.127 0.904 0.584 113,664 voicecraft/real_vs_synth – 12 0.436 0.130 0.901 0.581 113,568 e2_tts/synth_vs_real – 12 0.530 0.080 0.996 0.465 118,104 e2_tts/real_vs_synth – 12 0.530 0.079 0.992 0.466 118,344 H.10 E11 – Voice conversion Table 23: E11 – Voice conversion: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials freevc/spoof_vs_genuine – 12 0.225 0.096 0.620 0.832 60,000 real_real_baseline – 12 0.241 0.166 0.700 0.801 60,000 knn_vc/spoof_vs_genuine – 12 0.357 0.060 0.890 0.690 60,000 H.11 E13 – Age × codec Table 24: E13 – Age × codec: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials all__clean CV 12 0.108 0.119 0.462 0.933 120,000 young__clean CV 12 0.111 0.117 0.449 0.929 39,996 all__opus CV 12 0.112 0.118 0.471 0.931 120,000 young__opus CV 12 0.114 0.116 0.462 0.928 39,996 all__g711 CV 12 0.117 0.111 0.528 0.929 120,000 young__g711 CV 12 0.121 0.111 0.498 0.924 39,996 young__amr CV 12 0.123 0.109 0.537 0.924 39,996 all__amr CV 12 0.124 0.108 0.555 0.927 120,000 all__gsm CV 12 0.127 0.108 0.581 0.925 120,000 young__gsm CV 12 0.127 0.104 0.542 0.922 39,996 older__clean CV 12 0.130 0.129 0.420 0.907 39,996 older__opus CV 12 0.132 0.130 0.438 0.902 39,996 middle__clean CV 12 0.134 0.118 0.449 0.907 39,996 middle__opus CV 12 0.134 0.117 0.472 0.905 39,996 older__g711 CV 12 0.135 0.116 0.509 0.895 39,996 older__amr CV 12 0.141 0.110 0.542 0.893 39,996 older__gsm CV 12 0.143 0.092 0.545 0.899 39,996 middle__g711 CV 12 0.148 0.115 0.508 0.900 39,996 middle__amr CV 12 0.153 0.109 0.548 0.895 39,996 middle__gsm CV 12 0.158 0.110 0.582 0.890 39,996 H.12 E14 – Emotion × noise Table 25: E14 – Emotion × noise: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials neutral__clean__None ravdess 12 0.054 0.079 0.168 0.972 3,456 calm__gaussian__25 ravdess 12 0.067 0.099 0.236 0.955 3,456 calm__clean__None ravdess 12 0.072 0.088 0.226 0.964 3,456 neutral__gaussian__25 ravdess 12 0.072 0.107 0.257 0.953 3,456 fearful__clean__None ravdess 12 0.073 0.092 0.228 0.962 3,456 angry__clean__None ravdess 12 0.078 0.095 0.260 0.948 3,456 neutral__gaussian__15 ravdess 12 0.078 0.120 0.247 0.946 3,456 happy__clean__None ravdess 12 0.089 0.108 0.263 0.942 3,456 neutral__gaussian__5 ravdess 12 0.090 0.133 0.282 0.938 3,456 sad__clean__None ravdess 12 0.090 0.109 0.260 0.949 3,456 surprise__clean__None ravdess 12 0.093 0.117 0.292 0.945 3,456 disgust__clean__None ravdess 12 0.094 0.120 0.308 0.946 3,456 angry__gaussian__25 ravdess 12 0.094 0.120 0.308 0.926 3,456 fearful__gaussian__25 ravdess 12 0.095 0.124 0.267 0.937 3,456 happy__gaussian__25 ravdess 12 0.096 0.123 0.298 0.929 3,456 calm__gaussian__15 ravdess 12 0.097 0.136 0.300 0.933 3,456 calm__gaussian__5 ravdess 12 0.098 0.144 0.286 0.930 3,456 happy__gaussian__15 ravdess 12 0.105 0.146 0.324 0.917 3,456 sad__gaussian__25 ravdess 12 0.107 0.135 0.315 0.917 3,456 fearful__gaussian__15 ravdess 12 0.108 0.146 0.288 0.920 3,456 angry__gaussian__15 ravdess 12 0.112 0.153 0.339 0.912 3,456 sad__gaussian__5 ravdess 12 0.114 0.148 0.295 0.918 3,456 surprise__gaussian__25 ravdess 12 0.115 0.146 0.306 0.909 3,456 sad__gaussian__15 ravdess 12 0.115 0.156 0.316 0.911 3,456 fearful__gaussian__5 ravdess 12 0.119 0.160 0.313 0.911 3,456 disgust__gaussian__5 ravdess 12 0.120 0.148 0.336 0.916 3,456 surprise__gaussian__15 ravdess 12 0.120 0.160 0.341 0.908 3,456 disgust__gaussian__15 ravdess 12 0.122 0.158 0.337 0.904 3,456 happy__gaussian__5 ravdess 12 0.122 0.167 0.339 0.908 3,456 disgust__gaussian__25 ravdess 12 0.124 0.157 0.332 0.908 3,456 surprise__gaussian__5 ravdess 12 0.132 0.159 0.396 0.902 3,456 angry__gaussian__5 ravdess 12 0.136 0.174 0.359 0.897 3,456 fearful__environmental__25 ravdess 12 0.149 0.181 0.406 0.874 3,456 calm__environmental__25 ravdess 12 0.156 0.192 0.388 0.865 3,456 happy__environmental__25 ravdess 12 0.159 0.190 0.400 0.861 3,456 angry__environmental__25 ravdess 12 0.160 0.189 0.403 0.859 3,456 sad__environmental__25 ravdess 12 0.161 0.193 0.404 0.862 3,456 disgust__environmental__25 ravdess 12 0.165 0.199 0.410 0.859 3,456 neutral__environmental__25 ravdess 12 0.170 0.207 0.401 0.853 3,456 surprise__environmental__25 ravdess 12 0.172 0.205 0.411 0.850 3,456 fearful__environmental__15 ravdess 12 0.172 0.209 0.411 0.845 3,456 sad__environmental__15 ravdess 12 0.175 0.212 0.411 0.841 3,456 calm__environmental__15 ravdess 12 0.175 0.211 0.398 0.840 3,456 happy__environmental__15 ravdess 12 0.177 0.212 0.407 0.839 3,456 disgust__environmental__15 ravdess 12 0.178 0.209 0.407 0.836 3,456 neutral__environmental__15 ravdess 12 0.178 0.217 0.403 0.835 3,456 angry__environmental__15 ravdess 12 0.182 0.215 0.415 0.838 3,456 surprise__environmental__15 ravdess 12 0.188 0.222 0.457 0.827 3,456 calm__crosstalk__25 ravdess 12 0.194 0.215 0.462 0.825 3,456 sad__environmental__5 ravdess 12 0.194 0.227 0.439 0.812 3,456 fearful__environmental__5 ravdess 12 0.194 0.230 0.444 0.811 3,456 neutral__environmental__5 ravdess 12 0.195 0.233 0.411 0.812 3,456 angry__environmental__5 ravdess 12 0.196 0.226 0.443 0.812 3,456 calm__environmental__5 ravdess 12 0.196 0.228 0.435 0.810 3,456 happy__environmental__5 ravdess 12 0.201 0.238 0.428 0.808 3,456 sad__crosstalk__25 ravdess 12 0.203 0.218 0.447 0.815 3,456 disgust__environmental__5 ravdess 12 0.204 0.233 0.448 0.805 3,456 neutral__crosstalk__25 ravdess 12 0.205 0.230 0.451 0.812 3,456 surprise__environmental__5 ravdess 12 0.205 0.232 0.501 0.805 3,456 disgust__crosstalk__25 ravdess 12 0.209 0.225 0.483 0.811 3,456 fearful__crosstalk__25 ravdess 12 0.213 0.226 0.466 0.816 3,456 angry__crosstalk__25 ravdess 12 0.214 0.231 0.448 0.807 3,456 happy__crosstalk__25 ravdess 12 0.216 0.234 0.472 0.805 3,456 surprise__crosstalk__25 ravdess 12 0.229 0.231 0.506 0.794 3,456 calm__crosstalk__15 ravdess 12 0.239 0.243 0.520 0.776 3,456 neutral__crosstalk__15 ravdess 12 0.245 0.249 0.516 0.769 3,456 happy__crosstalk__15 ravdess 12 0.247 0.238 0.552 0.770 3,456 sad__crosstalk__15 ravdess 12 0.254 0.242 0.545 0.766 3,456 disgust__crosstalk__15 ravdess 12 0.255 0.228 0.561 0.770 3,456 fearful__crosstalk__15 ravdess 12 0.256 0.225 0.555 0.772 3,456 angry__crosstalk__15 ravdess 12 0.267 0.236 0.612 0.760 3,456 surprise__crosstalk__15 ravdess 12 0.276 0.215 0.661 0.754 3,456 neutral__crosstalk__5 ravdess 12 0.343 0.158 0.910 0.695 3,456 happy__crosstalk__5 ravdess 12 0.369 0.135 0.933 0.673 3,456 calm__crosstalk__5 ravdess 12 0.371 0.138 0.835 0.678 3,456 angry__crosstalk__5 ravdess 12 0.380 0.113 0.936 0.649 3,456 sad__crosstalk__5 ravdess 12 0.381 0.129 0.921 0.653 3,456 disgust__crosstalk__5 ravdess 12 0.387 0.119 0.969 0.644 3,456 surprise__crosstalk__5 ravdess 12 0.397 0.107 0.951 0.637 3,456 fearful__crosstalk__5 ravdess 12 0.401 0.110 0.933 0.642 3,456 H.13 E15 – Pathology and codec (descriptive) Table 26: E15 – Pathology and codec (descriptive): condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials EARS/healthy_clean ears 12 0.246 0.175 0.663 0.797 119,988 EARS/healthy_opus ears 12 0.259 0.160 0.679 0.787 29,997 EARS/healthy_amr ears 12 0.260 0.165 0.702 0.786 29,996 EARS/healthy_gsm ears 12 0.260 0.163 0.718 0.788 29,996 EARS/healthy_g711 ears 12 0.279 0.155 0.755 0.772 29,998 TORGO/dysarthria_moderate_clean torgo 12 0.303 0.106 0.932 0.752 37,411 TORGO/dysarthria_moderate_g711 torgo 12 0.304 0.100 0.880 0.758 9,376 TORGO/dysarthria_moderate_gsm torgo 12 0.305 0.113 0.921 0.744 9,367 TORGO/dysarthria_moderate_amr torgo 12 0.308 0.112 0.916 0.744 9,359 TORGO/dysarthria_severe_g711 torgo 12 0.309 0.121 0.946 0.744 9,645 TORGO/dysarthria_moderate_opus torgo 12 0.316 0.089 0.930 0.731 9,356 TORGO/dysarthria_severe_clean torgo 12 0.327 0.099 0.975 0.725 38,450 TORGO/dysarthria_severe_opus torgo 12 0.333 0.098 0.967 0.717 9,607 TORGO/dysarthria_severe_gsm torgo 12 0.334 0.099 0.964 0.717 9,611 TORGO/dysarthria_severe_amr torgo 12 0.337 0.097 0.951 0.717 9,622 TORGO/dysarthria_mild_clean torgo 12 0.339 0.096 0.972 0.708 38,187 TORGO/dysarthria_mild_opus torgo 12 0.342 0.088 0.954 0.706 9,521 TORGO/dysarthria_mild_gsm torgo 12 0.350 0.092 0.949 0.696 9,548 TORGO/dysarthria_mild_amr torgo 12 0.350 0.096 0.938 0.697 9,482 TORGO/dysarthria_mild_g711 torgo 12 0.351 0.099 0.951 0.697 9,550 H.14 E16 – Gender × noise Table 27: E16 – Gender × noise: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials CV/all_clean CV 12 0.109 0.118 0.446 0.932 60,000 CV/F_clean CV 12 0.110 0.113 0.432 0.932 60,000 CV/M_clean CV 12 0.117 0.116 0.466 0.927 60,000 CV/all_gaussian_25dB CV 12 0.131 0.135 0.453 0.912 13,332 CV/F_environmental_25dB CV 12 0.131 0.138 0.448 0.910 6,660 CV/all_gaussian_15dB CV 12 0.137 0.143 0.496 0.900 13,332 CV/F_gaussian_25dB CV 12 0.139 0.134 0.450 0.910 6,660 CV/F_gaussian_15dB CV 12 0.139 0.137 0.503 0.906 6,660 CV/all_environmental_25dB CV 12 0.140 0.144 0.437 0.901 13,332 CV/M_gaussian_25dB CV 12 0.141 0.149 0.410 0.899 6,660 CV/M_gaussian_15dB CV 12 0.146 0.132 0.509 0.897 6,660 CV/M_environmental_25dB CV 12 0.147 0.142 0.457 0.896 6,660 CV/all_environmental_15dB CV 12 0.154 0.154 0.522 0.889 13,332 CV/F_gaussian_5dB CV 12 0.159 0.148 0.579 0.884 6,660 CV/M_crosstalk_25dB CV 12 0.162 0.146 0.430 0.885 6,660 CV/all_crosstalk_25dB CV 12 0.163 0.154 0.495 0.880 13,332 CV/F_crosstalk_25dB CV 12 0.167 0.145 0.444 0.882 6,660 CV/all_gaussian_5dB CV 12 0.167 0.160 0.574 0.875 13,332 CV/F_environmental_15dB CV 12 0.171 0.162 0.527 0.876 6,660 CV/M_gaussian_5dB CV 12 0.180 0.153 0.565 0.866 6,660 CV/M_environmental_15dB CV 12 0.187 0.169 0.523 0.856 6,660 CV/all_environmental_5dB CV 12 0.193 0.182 0.554 0.836 13,332 CV/M_environmental_5dB CV 12 0.200 0.171 0.535 0.847 6,660 CV/F_environmental_5dB CV 12 0.201 0.181 0.525 0.834 6,660 CV/all_crosstalk_15dB CV 12 0.227 0.165 0.604 0.821 13,332 CV/M_crosstalk_15dB CV 12 0.236 0.188 0.551 0.803 6,660 CV/F_crosstalk_15dB CV 12 0.270 0.172 0.616 0.784 6,660 CV/all_crosstalk_5dB CV 12 0.333 0.121 0.928 0.720 13,332 CV/M_crosstalk_5dB CV 12 0.344 0.104 0.855 0.712 6,660 CV/F_crosstalk_5dB CV 12 0.347 0.108 0.854 0.714 6,660 H.15 E17 – Temporal drift × codec Table 28: E17 – Temporal drift × codec: condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials VoxCeleb1/gap_0_clean voxceleb1 12 0.168 0.186 0.502 0.862 120,000 VoxCeleb1/gap_0_opus voxceleb1 12 0.169 0.180 0.496 0.862 30,000 VoxCeleb1/gap_5yr_opus voxceleb1 12 0.176 0.183 0.523 0.860 9,984 VoxCeleb1/gap_1yr_clean voxceleb1 12 0.176 0.189 0.516 0.853 39,984 VoxCeleb1/gap_5yr_clean voxceleb1 12 0.176 0.187 0.516 0.856 39,984 VoxCeleb1/gap_3yr_opus voxceleb1 12 0.178 0.181 0.496 0.854 9,984 VoxCeleb1/gap_0_g711 voxceleb1 12 0.179 0.177 0.567 0.856 30,000 VoxCeleb1/gap_3yr_clean voxceleb1 12 0.180 0.192 0.527 0.849 39,984 VoxCeleb1/gap_1yr_opus voxceleb1 12 0.183 0.197 0.499 0.847 9,984 VoxCeleb1/gap_0_amr voxceleb1 12 0.185 0.177 0.608 0.853 30,000 VoxCeleb1/gap_5yr_g711 voxceleb1 12 0.190 0.182 0.602 0.850 9,984 VoxCeleb1/gap_3yr_g711 voxceleb1 12 0.192 0.177 0.572 0.846 9,984 VoxCeleb1/gap_0_gsm voxceleb1 12 0.195 0.182 0.645 0.842 30,000 VoxCeleb1/gap_1yr_g711 voxceleb1 12 0.196 0.195 0.542 0.839 9,984 VoxCeleb1/gap_3yr_amr voxceleb1 12 0.198 0.174 0.588 0.839 9,984 VoxCeleb1/gap_5yr_amr voxceleb1 12 0.199 0.180 0.590 0.842 9,984 VoxCeleb1/gap_1yr_amr voxceleb1 12 0.202 0.193 0.576 0.834 9,984 VoxCeleb1/gap_1yr_gsm voxceleb1 12 0.206 0.193 0.613 0.828 9,984 VoxCeleb1/gap_5yr_gsm voxceleb1 12 0.209 0.180 0.609 0.837 9,984 VoxCeleb1/gap_3yr_gsm voxceleb1 12 0.209 0.174 0.620 0.835 9,984 H.16 E18 – Accent and codec (descriptive) Table 29: E18 – Accent and codec (descriptive): condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials CV_EN/native_gsm CV 12 0.105 0.117 0.388 0.936 15,000 CV_EN/native_amr CV 12 0.106 0.119 0.381 0.934 15,000 CV_EN/native_clean CV 12 0.109 0.119 0.438 0.933 60,000 CV_EN/native_opus CV 12 0.113 0.120 0.424 0.930 15,000 CV_EN/native_g711 CV 12 0.115 0.111 0.456 0.929 15,000 L2-ARCTIC/accent_gsm L2-ARCTIC 12 0.175 0.171 0.649 0.859 14,951 L2-ARCTIC/accent_amr L2-ARCTIC 12 0.177 0.174 0.551 0.854 14,951 L2-ARCTIC/accent_clean L2-ARCTIC 12 0.179 0.169 0.650 0.856 59,828 L2-ARCTIC/accent_opus L2-ARCTIC 12 0.180 0.166 0.673 0.853 14,957 L2-ARCTIC/accent_g711 L2-ARCTIC 12 0.188 0.168 0.682 0.850 14,961 H.17 E20 – Age × codec × drift (age proxied) Table 30: E20 – Age × codec × drift (age proxied): condition-level means over the contributing encoder rows. SD is over contributing encoder rows, not a confidence interval. Condition Corpus Enc. EER SD minDCF AUC Trials VoxCeleb1/age_middle_clean voxceleb1 12 0.163 0.180 0.463 0.868 15,000 VoxCeleb1/AC_older_opus voxceleb1 12 0.163 0.178 0.488 0.869 7,500 VoxCeleb1_3way/older_gap5yr_opus voxceleb1 12 0.166 0.178 0.491 0.862 3,312 VoxCeleb1/gap0_opus voxceleb1 12 0.167 0.180 0.471 0.867 9,996 VoxCeleb1/AC_middle_opus voxceleb1 12 0.167 0.177 0.469 0.866 7,500 VoxCeleb1/age_older_clean voxceleb1 12 0.167 0.183 0.473 0.865 15,000 VoxCeleb1/clean voxceleb1 12 0.168 0.184 0.488 0.863 30,000 VoxCeleb1/BC_older_gap5yr voxceleb1 12 0.168 0.181 0.476 0.864 7,488 VoxCeleb1/age_young_clean voxceleb1 12 0.168 0.183 0.467 0.865 15,000 VoxCeleb1/BC_middle_gap5yr voxceleb1 12 0.169 0.186 0.459 0.860 7,488 VoxCeleb1/AC_young_opus voxceleb1 12 0.170 0.183 0.482 0.864 7,500 VoxCeleb1/gap_5yr_clean voxceleb1 12 0.171 0.180 0.529 0.863 9,984 VoxCeleb1_3way/middle_gap5yr_opus voxceleb1 12 0.173 0.192 0.437 0.863 3,312 VoxCeleb1/AC_older_g711 voxceleb1 12 0.174 0.170 0.512 0.867 7,500 VoxCeleb1/gap_3yr_clean voxceleb1 12 0.175 0.181 0.485 0.857 9,984 VoxCeleb1_3way/middle_gap5yr_g711 voxceleb1 12 0.177 0.183 0.502 0.858 3,312 VoxCeleb1/AC_young_g711 voxceleb1 12 0.177 0.170 0.539 0.861 7,500 VoxCeleb1/gap_5yr_opus voxceleb1 12 0.178 0.182 0.457 0.860 3,312 VoxCeleb1/BC_older_gap3yr voxceleb1 12 0.179 0.191 0.485 0.853 7,488 VoxCeleb1/BC_young_gap3yr voxceleb1 12 0.180 0.189 0.483 0.851 7,488 VoxCeleb1/gap0_g711 voxceleb1 12 0.183 0.183 0.541 0.854 9,996 VoxCeleb1_3way/older_gap5yr_g711 voxceleb1 12 0.183 0.169 0.599 0.859 3,312 VoxCeleb1/AC_middle_g711 voxceleb1 12 0.184 0.181 0.534 0.852 7,500 VoxCeleb1_3way/young_gap3yr_opus voxceleb1 12 0.184 0.193 0.515 0.845 3,312 VoxCeleb1_3way/middle_gap5yr_gsm voxceleb1 12 0.184 0.175 0.525 0.853 3,312 VoxCeleb1/gap_5yr_g711 voxceleb1 12 0.185 0.183 0.521 0.856 3,312 VoxCeleb1/BC_middle_gap3yr voxceleb1 12 0.187 0.195 0.505 0.844 7,488 VoxCeleb1/gap_3yr_opus voxceleb1 12 0.187 0.200 0.478 0.845 3,312 VoxCeleb1_3way/young_gap5yr_opus voxceleb1 12 0.188 0.186 0.493 0.845 3,312 VoxCeleb1/gap_3yr_g711 voxceleb1 12 0.188 0.185 0.597 0.841 3,312 VoxCeleb1/BC_young_gap5yr voxceleb1 12 0.188 0.199 0.486 0.842 7,488 VoxCeleb1/AC_older_gsm voxceleb1 12 0.189 0.180 0.574 0.852 7,500 VoxCeleb1/gap0_gsm voxceleb1 12 0.190 0.174 0.597 0.845 9,996 VoxCeleb1_3way/older_gap5yr_gsm voxceleb1 12 0.193 0.173 0.632 0.850 3,312 VoxCeleb1_3way/young_gap3yr_g711 voxceleb1 12 0.194 0.183 0.502 0.841 3,312 VoxCeleb1_3way/middle_gap3yr_opus voxceleb1 12 0.194 0.200 0.479 0.835 3,312 VoxCeleb1/AC_middle_gsm voxceleb1 12 0.197 0.176 0.585 0.843 7,500 VoxCeleb1/AC_young_gsm voxceleb1 12 0.197 0.177 0.598 0.843 7,500 VoxCeleb1/gap_5yr_gsm voxceleb1 12 0.199 0.173 0.557 0.847 3,312 VoxCeleb1_3way/older_gap3yr_opus voxceleb1 12 0.199 0.192 0.542 0.836 3,312 VoxCeleb1_3way/young_gap5yr_g711 voxceleb1 12 0.204 0.185 0.556 0.833 3,312 VoxCeleb1/gap_3yr_gsm voxceleb1 12 0.205 0.173 0.572 0.831 3,312 VoxCeleb1_3way/middle_gap3yr_g711 voxceleb1 12 0.208 0.194 0.535 0.829 3,312 VoxCeleb1_3way/young_gap3yr_gsm voxceleb1 12 0.208 0.183 0.560 0.834 3,312 VoxCeleb1_3way/older_gap3yr_g711 voxceleb1 12 0.211 0.187 0.592 0.829 3,312 VoxCeleb1_3way/young_gap5yr_gsm voxceleb1 12 0.214 0.171 0.601 0.828 3,312 VoxCeleb1_3way/older_gap3yr_gsm voxceleb1 12 0.215 0.183 0.626 0.820 3,312 VoxCeleb1_3way/middle_gap3yr_gsm voxceleb1 12 0.230 0.194 0.601 0.814 3,312