Speaker verification error on the VoxCeleb1 test fell from 8.8% in 2017 (Nagrani et al., VoxCeleb paper 2017) to 0.12% in a 2025 paper (Li, Cheng and Li, w2v-BERT 2.0 paper 2025), a 73-fold drop (8.8 / 0.12). Speaker verification is the “is this the same person?” check, and equal error rate (EER) is its standard score. That record comes from a fixed list of celebrity speech taken from YouTube. The same field reports 9.30% when crowd chatter is as loud as the voice (Gu et al., Noise-Conditioned Mixture-of-Experts paper 2025) and 10.419% when one person is compared across a large age gap (Qin et al., Interspeech 2022). We aggregated data from the VoxSRC challenge organizers, the NIST Speaker Recognition Evaluation, Interspeech and Odyssey papers, the VoxBlink2 and CN-Celeb corpus teams, and FAccT fairness research. For the commercial side of this technology, see our voice biometrics statistics, and for the “who spoke when” problem that sits next to it, see our speaker diarization error rate roundup.
TL;DR
- Lowest VoxCeleb1-O EER we found in a primary paper: 0.12%, from a model with about 600 million parameters (Li, Cheng and Li, w2v-BERT 2.0 paper 2025).
- A corpus of 111,284 speakers helped push a single ResNet to 0.17% EER and 0.006 minDCF (Lin et al., VoxBlink2, Interspeech 2024).
- ReDimNet2-B6 reaches 0.29% EER with 12.3M parameters (Yakovlev and Okhotnikov, ReDimNet2 paper 2026).
- Track 2 first-place EER on the fixed 2019 VoxSRC test set fell from 1.26% in 2019 to 0.47% in 2023 (Huh et al., VoxSRC retrospective 2024).
- A 2-second trial raises ERes2NetV2 from 0.61% to 1.48% EER (Chen et al., ERes2NetV2 paper 2024).
- Babble at 0 dB raises a ResNet34 baseline from 1.98% to 9.30% EER (Gu et al. 2025).
- Pairs with large age gaps raise a baseline from 1.939% to 10.419% EER (Qin et al., Interspeech 2022).
- A 40-year gap took male EER to 51.61% in a Finnish corpus, past the 50% coin-flip line (Singh et al., Interspeech 2023).
- An x-vector system trained on VoxCeleb scored 3.48% on SITW and 16.59% on CN-Celeb.E (Li et al., CN-Celeb paper 2020).
- The SRE24 telephone challenge leaderboard was led at 2.37% EER (NIST SRE24-CTS Challenge leaderboard, October 24, 2024).
- Women were 38% of speakers but 29.5% of utterances in the VoxCeleb 1 test set (Hutiri and Ding, FAccT 2022).
1. The Yardstick: What EER Measures and What VoxCeleb Contains
A verification system receives two recordings, called a “trial”, and decides whether they come from the same speaker. Raising the acceptance threshold rejects more real speakers, and lowering it accepts more impostors. EER is the point where those two error rates match. The companion metric, minDCF (minimum detection cost), weights the same two errors for a chosen operating point, and lower is also better. Neither metric covers synthetic or replayed voices: the “impostors” here are simply other real people.
The classic test list is small: 37,720 trials from 40 speakers (Gu et al., Noise-Conditioned Mixture-of-Experts paper 2025). VoxCeleb1-O, the name for that list, is easy compared with its harder siblings: one model’s error quadruples when moved from Vox1-O to Vox1-H (0.17% to 0.68%). Training data, by contrast, has grown enormously, and section 3 shows why that matters more than most architecture changes.
| Metric | Value | Source |
|---|---|---|
| VoxCeleb1 size: speakers, utterances, videos, hours, nationalities | 1,251 / 153,516 / 22,496 / 352 / 36 | Chung et al., VoxCeleb2 paper 2018 |
| VoxCeleb2 training portion: speakers, utterances, hours, average seconds per utterance | 5,994 / 1,092,009 / 2,442 / 7.8 | Lin et al., VoxBlink2, Interspeech 2024 |
| VoxBlink2 size: speakers, utterances, hours, average seconds per utterance | 111,284 / 9,904,382 / 16,672 / 6.0 | Lin et al., VoxBlink2, Interspeech 2024 |
| CN-Celeb size: speakers, recording genres, hours in CN-Celeb1 plus CN-Celeb2 | 3,000 / 11 / 274 plus 1,090 | Li et al., CN-Celeb paper 2020 |
| VoxCeleb1-O test list: trials, speakers | 37,720 / 40 | Gu et al., Noise-Conditioned Mixture-of-Experts paper 2025 |
| EER of one model on Vox1-O (easy list) vs Vox1-E (extended list) vs Vox1-H (hard list) | 0.17% vs 0.37% vs 0.68% | Lin et al., VoxBlink2, Interspeech 2024 |
Most recent available data for the original corpus figures: VoxCeleb2 paper, 2018.
2. The VoxSRC Challenge Record, 2019 to 2023
The VoxCeleb Speaker Recognition Challenge ran every year from 2019 to 2023, and its best design choice was keeping the 2019 test set inside every later test set. That gives the field one fixed ruler. On it, Track 1 first-place EER fell from 1.42% in 2019 to 0.57% in 2021, then stopped falling. The organizers say winners have been below 1% since 2020, “which we consider as performance saturation”, and each year’s new test set stayed harder: Track 1 winners scored 3.73%, 1.85%, 1.49% and 1.59% on their own sets from 2020 to 2023, as the test grew from 208,008 pairs and 745 speakers in 2019 to 825,437 pairs and 2,532 speakers in 2023 (Huh et al., VoxSRC retrospective 2024).
Saturation has a measurable floor. On that 2019 set, 95% bootstrap confidence intervals were 18.6% to 31.2% as wide as the EER itself, so small gaps between winners may be noise. Of the 232 pairs that all 2022 and 2023 winners got wrong, 16 were same-speaker pairs, and 12 of those 16 were label errors in the test list. The hard cases are real too: the 2022 Track 1 winner scored 0.90% on standard pairs and 2.07% when every pair was a hard one (the same speaker at different ages, or different speakers in the same recording environment).
| Metric | Value | Source |
|---|---|---|
| Track 1 (closed training data) first-place EER on the fixed 2019 test set, 2019 / 2020 / 2021 / 2022 / 2023 | 1.42% / 0.83% / 0.57% / 0.90% / 0.58% | Huh et al., VoxSRC retrospective 2024 |
| Track 2 (open training data) first-place EER on the fixed 2019 test set, 2019 / 2020 / 2021 / 2022 / 2023 | 1.26% / 0.80% / 0.57% / 0.69% / 0.47% | Huh et al., VoxSRC retrospective 2024 |
| UNISOUND 2023 entry, first in Track 1 by the team’s report: EER, minDCF on the 2023 test set | 1.5880%, 0.0855 | Zheng et al., UNISOUND system report 2023 |
| ID R&D 2023 entry, first on the Track 2 public leaderboard: EER, minDCF at P_target 0.05 | 1.30%, 0.0762 | Torgashov et al., ID R&D system report 2023 |
| ChinaTelecom 2023 Track 1 fused entry, self-reported: EER, minDCF | 1.980%, 0.1066 | Du et al., ChinaTelecom system report 2023 |
Context: the final edition was held in 2023, so most recent available data for this challenge is the 2024 retrospective. EER was often not the ranking metric, which is why a second-place team sometimes shows a lower EER than the winner.
3. The VoxCeleb1-O Record, From 8.8% to 0.12%
The trial count makes the improvement concrete. Both error rates equal the EER at the operating point, so total mistakes are about EER times trials: 8.8% of 37,720 is roughly 3,319 wrong decisions, and 0.12% is roughly 45. A 2025 system makes about 45 mistakes where the 2017 baseline made about 3,319 (Li, Cheng and Li, w2v-BERT 2.0 paper 2025). At that level, a handful of mislabeled trials can decide a record.
Data now moves the score more than architecture. W2V-BERT 2.0 is listed at 0.38% when trained on VoxCeleb2 alone and 0.14% with VoxBlink2 added, the same 587M-parameter model in both rows (Yakovlev and Okhotnikov, ReDimNet2 paper 2026). Recipes matter as well: ResNet34 went from 1.31% in its cited original to 0.723% in the WeSpeaker toolkit re-implementation. These same embeddings also serve as judges in voice conversion research, which our voice conversion speaker similarity roundup examines.
| Metric | Value | Source |
|---|---|---|
| 2017: i-vector plus PLDA baseline on the VoxCeleb1 test, EER and minimum detection cost | 8.8%, 0.73 | Nagrani et al., VoxCeleb paper 2017 |
| 2018: ResNet-50 trained on VoxCeleb2, EER and minimum detection cost | 3.95%, 0.429 | Chung et al., VoxCeleb2 paper 2018 |
| 2020: ECAPA-TDNN with 1,024 channels, EER and minDCF | 0.87%, 0.1066 | Desplanques et al., Interspeech 2020 |
| 2022: ResNet293 in the WeSpeaker toolkit, EER and minDCF | 0.447%, 0.043 | Wang et al., WeSpeaker paper 2022 |
| 2024: SimAM-ResNet293 pre-trained on VoxBlink2, with quality calibration, EER and minDCF | 0.17%, 0.006 | Lin et al., VoxBlink2, Interspeech 2024 |
| 2025: w2v-BERT 2.0 on Vox1-O vs Vox1-H, EER | 0.12% vs 0.55% | Li, Cheng and Li, w2v-BERT 2.0 paper 2025 |
| 2026: ReDimNet2-B6, EER, parameters, GMACs (billions of multiply-accumulate operations) for 2 seconds of input | 0.29%, 12.3M, 13.05 | Yakovlev and Okhotnikov, ReDimNet2 paper 2026 |
| 2026: ReDimNet2-B0, the smallest model in the family, EER and parameters | 1.04%, 1.1M | Yakovlev and Okhotnikov, ReDimNet2 paper 2026 |
Outlier note: the w2v-BERT 2.0 authors report that pruning cut model size by 80% with only a 0.04% EER degradation, so the largest models may not need to stay large.
4. Short Clips and Noisy Audio
Cutting a test clip to 2 seconds more than doubles the error of a leading model, from 0.61% to 1.48% (Chen et al., ERes2NetV2 paper 2024). That matters for any product built on a brief spoken phrase, since the sub-1% headline belongs to full-length clips. The reverse also holds: with VoxSRC pairs longer than 8 seconds, the 2023 Track 2 winner almost ran out of errors.
Noise follows a steep curve, not a straight line. “Babble” means background crowd chatter, and SNR (signal-to-noise ratio) measures the voice against the noise, so 0 dB means both are equally loud. A ResNet34 baseline barely moves at 20 dB, then its error climbs more than fourfold at 0 dB (Gu et al., Noise-Conditioned Mixture-of-Experts paper 2025). Crowd chatter hurt more than music, likely because it is the same kind of signal the model is trained to identify.
| Metric | Value | Source |
|---|---|---|
| ERes2NetV2 EER on VoxCeleb1-O: full-length vs 3-second vs 2-second trials | 0.61% vs 0.98% vs 1.48% | Chen et al., ERes2NetV2 paper 2024 |
| ECAPA-TDNN EER on VoxCeleb1-O: 3-second vs 2-second trials | 1.27% vs 1.95% | Chen et al., ERes2NetV2 paper 2024 |
| ResNet34 EER on VoxCeleb1-O: 3-second vs 2-second trials | 1.47% vs 2.12% | Chen et al., ERes2NetV2 paper 2024 |
| 2023 VoxSRC Track 2 winner EER on the 2019 test set: all pairs vs pairs over 6 seconds vs pairs over 8 seconds | 0.47% vs 0.26% vs 0.03% | Huh et al., VoxSRC retrospective 2024 |
| SRE24 test segments where systems struggled most, and the point above which gains shrink | 4.5 to 13 seconds; smaller gains beyond 22.6 seconds | Singer et al., Odyssey 2026 |
| ResNet34 baseline EER on the VoxCeleb1 test: clean vs babble at 20 dB vs babble at 0 dB | 1.98% vs 2.18% vs 9.30% | Gu et al., Noise-Conditioned Mixture-of-Experts paper 2025 |
| Same baseline at 0 dB: music vs general noise | 5.82% vs 7.3% | Gu et al., Noise-Conditioned Mixture-of-Experts paper 2025 |
| Same baseline at 0 dB with Nonspeech100 noise, and the noise-conditioned model at 0 dB babble | 10.17%; 8.10% | Gu et al., Noise-Conditioned Mixture-of-Experts paper 2025 |
Context: noise results do not transfer between papers. A University of Seoul study measured ECAPA-TDNN at 3.46% clean and 13.01% with babble at 0 dB (Kim et al., PAS paper 2023), under a different training setup, so compare the rise inside each study rather than the absolute values.
5. Age, Corpus and Domain Mismatch
A record on VoxCeleb1-O says little about a voice recorded years apart or in another setting. Choosing same-speaker pairs with large age gaps raised a baseline from 1.939% to 10.419% EER (Qin et al., Interspeech 2022). A Finnish longitudinal corpus shows the long end of that curve: at a 40-year gap, male EER of 51.61% is worse than guessing (Singh et al., Interspeech 2023). That corpus is small, with 109 speakers after preprocessing, and VoxCeleb ages are estimated rather than recorded, so read the direction with more confidence than the exact values.
Domain shift is the larger commercial risk, because it arrives silently. The VoxCeleb-trained x-vector in the CN-Celeb paper saw its error rise nearly fivefold on Chinese multi-genre audio (16.59 / 3.48 = 4.8). Retraining on the new domain helps but does not close the gap: the same ECAPA-TDNN recipe still showed more than nine times the error on CN-Celeb that it showed on VoxCeleb (8.877 / 0.936 = 9.5).
| Metric | Value | Source |
|---|---|---|
| Baseline EER on the Vox-H test vs Vox-CA20, a test of pairs with large age gaps | 1.939% vs 10.419% | Qin et al., Interspeech 2022 |
| ECAPA-TDNN EER on VoxCeleb1-Age-Enriched (USA) at 3 / 6 / 9 years of age gap, male speakers | 2.31% / 2.84% / 3.63% | Singh et al., Interspeech 2023 |
| Same test at 3 / 6 / 9 years of age gap, female speakers | 2.70% / 3.12% / 3.87% | Singh et al., Interspeech 2023 |
| Finnish longitudinal corpus (LCFSH) EER at 0 / 20 / 40 years of age gap, male speakers | 5.12% / 18.68% / 51.61% | Singh et al., Interspeech 2023 |
| Same corpus at 0 / 20 / 40 years of age gap, female speakers | 10.04% / 21.67% / 40.61% | Singh et al., Interspeech 2023 |
| VoxCeleb-trained x-vector EER on the SITW test set vs CN-Celeb.E (an evaluation set from CN-Celeb) | 3.48% vs 16.59% | Li et al., CN-Celeb paper 2020 |
| ECAPA-TDNN with AM-Softmax loss: VoxCeleb1 test vs CN-Celeb evaluation, each trained on its own domain | 0.936% vs 8.877% | Wang et al., DASA paper 2023 |
| VoxSRC 2023 domain adaptation track on CNCeleb: baseline vs first place, EER | 13.11% vs 4.95% | Huh et al., VoxSRC retrospective 2024 |
Most recent available data for long-term aging: Singh et al., Interspeech 2023.
6. NIST SRE: The Harder Benchmark Without the Headline Numbers
NIST has run these evaluations since 1996, and they look nothing like YouTube interviews. SRE24 used a new corpus collected outside North America, telephone speech and audio from video, and Tunisian Arabic, English and French. The SRE24 telephone challenge leader sat at 2.37% EER, about 20 times the VoxCeleb1-O record (2.37 / 0.12 = 19.8) (NIST SRE24 site), though the two tests differ too much for that ratio to measure anything but difficulty.
The analysis of 91 submitted SRE24 systems found matched telephone trials consistently beat matched audio-from-video trials, and source type mattered more than language mismatch. Three enrollment segments generally beat one, even with equal total speech, and most systems showed limited sex differences on same-sex trials (Singer et al., Odyssey 2026). Unrestricted training data delivered significant gains, and single models competed with fusions (Greenberg et al., Interspeech 2025). Numbers in the table below show the other lesson: a system that cannot set its own threshold is not ready, however good its minimum cost.
| Metric | Value | Source |
|---|---|---|
| SRE19 telephone challenge: organizations, teams, valid system outputs | 67 / 51 / 1,347 | Sadjadi et al., NIST SRE19 CTS Challenge, Odyssey 2020 |
| SRE21 audio test: target trials vs non-target trials | 132,038 vs 5,899,731 | Sadjadi et al., NIST SRE21 2022 |
| SRE24: teams, organizations, submitted systems | 11 / 26 / 91 | Singer et al., Odyssey 2026 |
| SRE24-CTS leaderboard first row on October 24, 2024: EER, minimum cost, actual cost | 2.37%, 0.102, 0.106 | NIST SRE24-CTS Challenge leaderboard 2024 |
| Another entry on that leaderboard: EER, minimum cost, actual cost (a trivial system scores 1) | 3.00%, 0.080, 1.000 | NIST SRE24-CTS Challenge leaderboard 2024 |
| HCCL submitted SRE21 system, development vs evaluation set: EER | 2.27% vs 3.05% | Li et al., HCCL system, Interspeech 2022 |
| Same system, development vs evaluation set: actual cost | 0.170 vs 0.560 | Li et al., HCCL system, Interspeech 2022 |
Context: the SRE24 overview and analysis papers present results as charts, without printed EER values, so the leaderboard is the numeric window into this cycle. Most recent available data for official numeric results: NIST SRE24-CTS Challenge leaderboard, 2024.
7. Fairness: Who the Benchmark Measures
A single benchmark score can hide one group’s false positive rate running 13 times the overall rate. At the overall minimum-cost threshold, Indian female speakers had a false positive rate 13.0387 times the overall rate in a study of the VoxCeleb Speaker Recognition Challenge workflow (Hutiri and Ding, FAccT 2022). The composition of the data helps explain it: the VoxCeleb 1 test set is 62.5% US speakers, and women contribute 38% of speakers but 29.5% of utterances.
The direction of the gender gap is not fixed, which is the useful finding. VoxSRC winners did better on men, an Amazon study did better on women, and a controlled experiment shows the training mix can create a gap of its own. The same pattern shows up in word-level speech tools, as our speech recognition accent bias roundup documents.
| Metric | Value | Source |
|---|---|---|
| VoxCeleb 1 test set: US share of speakers; female share of speakers; female share of utterances | 62.5%; 38%; 29.5% | Hutiri and Ding, FAccT 2022 |
| VoxCeleb 2 training set: male share; US share | 61%; 29% | Hutiri and Ding, FAccT 2022 |
| Indian female speakers vs overall rate: false positive ratio; false negative ratio | 13.0387 times; 1.2497 times | Hutiri and Ding, FAccT 2022 |
| 2023 VoxSRC winners on the 2019 test set, male vs female EER: Track 1; Track 2 | 0.57% vs 0.69%; 0.43% vs 0.59% | Huh et al., VoxSRC retrospective 2024 |
| Adversarial reweighting system, male vs female EER | 1.25% vs 0.67% | Jin et al., Interspeech 2022 |
| Same study, gap between gender groups before vs after reweighting | 0.70% vs 0.58% | Jin et al., Interspeech 2022 |
| Same study, EER for US vs UK vs all other nationalities | 1.04% vs 0.76% vs 1.22% | Jin et al., Interspeech 2022 |
| ResNet-34 baseline (quarter-channel) trained on a 9:1 female-to-male mix, female vs male EER | 3.52% vs 7.22% | Shen et al., ICASSP 2022 |
Most recent available data for subgroup error ratios: Hutiri and Ding, FAccT 2022. Separately, Estevez and Ferrer found performance “markedly degraded” for groups underrepresented in training, and that simple data balancing reduced the bias without hurting majority groups (Study on the Fairness of Speaker Verification Systems).
Summary: Speaker Verification VoxCeleb Equal Error Rate by the Numbers
| Metric | Value | Source |
|---|---|---|
| Lowest VoxCeleb1-O EER found in a primary paper | 0.12% | Li, Cheng and Li, w2v-BERT 2.0 paper 2025 |
| Same system on the hard Vox1-H list | 0.55% | Li, Cheng and Li, w2v-BERT 2.0 paper 2025 |
| SimAM-ResNet293 on VoxCeleb1-O with VoxBlink2 pre-training and quality calibration, EER and minDCF | 0.17%, 0.006 | Lin et al., VoxBlink2, Interspeech 2024 |
| VoxBlink2 speakers vs VoxCeleb2 speakers | 111,284 vs 5,994 | Lin et al., VoxBlink2, Interspeech 2024 |
| ReDimNet2-B6 EER and parameters | 0.29%, 12.3M | Yakovlev and Okhotnikov, ReDimNet2 paper 2026 |
| ECAPA-TDNN with 1,024 channels, EER | 0.87% | Desplanques et al., Interspeech 2020 |
| VoxSRC Track 2 first-place EER on the fixed 2019 test set, 2019 vs 2023 | 1.26% vs 0.47% | Huh et al., VoxSRC retrospective 2024 |
| ERes2NetV2 EER on full-length vs 2-second trials | 0.61% vs 1.48% | Chen et al., ERes2NetV2 paper 2024 |
| ResNet34 baseline EER, clean vs babble at 0 dB | 1.98% vs 9.30% | Gu et al., Noise-Conditioned Mixture-of-Experts paper 2025 |
| Baseline EER on Vox-H vs Vox-CA20 (large age gaps) | 1.939% vs 10.419% | Qin et al., Interspeech 2022 |
| Male EER at a 0-year vs 40-year age gap, Finnish corpus | 5.12% vs 51.61% | Singh et al., Interspeech 2023 |
| VoxCeleb-trained x-vector EER on SITW vs CN-Celeb.E | 3.48% vs 16.59% | Li et al., CN-Celeb paper 2020 |
| ECAPA-TDNN EER on VoxCeleb1 vs CN-Celeb evaluation, each trained on its own domain | 0.936% vs 8.877% | Wang et al., DASA paper 2023 |
| SRE24-CTS leaderboard first row, EER | 2.37% | NIST SRE24-CTS Challenge leaderboard 2024 |
| HCCL SRE21 actual cost, development vs evaluation set | 0.170 vs 0.560 | Li et al., HCCL system, Interspeech 2022 |
| VoxCeleb 1 test set, women’s share of speakers vs utterances | 38% vs 29.5% | Hutiri and Ding, FAccT 2022 |
| Indian female false positive rate vs overall rate | 13.0387 times | Hutiri and Ding, FAccT 2022 |
| Adversarial reweighting system, male vs female EER | 1.25% vs 0.67% | Jin et al., Interspeech 2022 |
Methodology and Sources
Every figure was read in a primary paper, challenge report or NIST page during this research. EER and cost numbers are quoted as each source prints them. Ratios and wrong-decision counts are our own arithmetic, with the math shown. No VoxBooster data is used.
- Huh et al., The VoxCeleb Speaker Recognition Challenge: A Retrospective, 2024
- Nagrani, Chung and Zisserman, VoxCeleb: a large-scale speaker identification dataset, Interspeech 2017
- Chung, Nagrani and Zisserman, VoxCeleb2: Deep Speaker Recognition, Interspeech 2018
- Desplanques, Thienpondt and Demuynck, ECAPA-TDNN, Interspeech 2020
- Wang et al., WeSpeaker toolkit paper, 2022
- Lin et al., VoxBlink2, Interspeech 2024
- Li, Cheng and Li, w2v-BERT 2.0 and structured pruning, 2025
- Yakovlev and Okhotnikov, ReDimNet2, 2026
- Zheng et al., UNISOUND system for VoxSRC 2023
- Torgashov et al., ID R&D system for VoxSRC 2023
- Du, Fang and Li, ChinaTelecom system for VoxSRC 2023
- Chen et al., ERes2NetV2, 2024
- Gu et al., Noise-Conditioned Mixture-of-Experts Framework, 2025
- Kim et al., PAS augmentation paper, 2023
- Qin et al., Cross-Age Speaker Verification, Interspeech 2022
- Singh, Sahidullah and Kinnunen, Speaker Verification Across Ages, Interspeech 2023
- Li et al., CN-Celeb: multi-genre speaker recognition, 2020
- Wang et al., DASA paper, 2023
- NIST Speaker Recognition Evaluation site and SRE24-CTS Challenge leaderboard
- Greenberg et al., The 2024 NIST Speaker Recognition Evaluation, Interspeech 2025
- Singer et al., Analysis of the NIST 2024 Speaker Recognition Evaluation, Odyssey 2026
- Sadjadi et al., The 2021 NIST Speaker Recognition Evaluation, 2022
- Sadjadi et al., The 2019 NIST SRE CTS Challenge, Odyssey 2020
- Li et al., The HCCL System for the NIST SRE21, Interspeech 2022
- Hutiri and Ding, Bias in Automated Speaker Recognition, FAccT 2022
- Jin et al., Adversarial Reweighting for Speaker Verification Fairness, Interspeech 2022
- Shen et al., Group-adapted Fusion Network, ICASSP 2022
- Estevez and Ferrer, Study on the Fairness of Speaker Verification Systems Across Accent and Gender Groups
- Qu et al., Fair-Gate, 2026
- Data watch: VoxCeleb1-O records are not perfectly comparable, because some papers use a cleaned version of the protocol, and some add score normalization or quality calibration. The ECAPA-TDNN figure alone varies by paper: 0.87% in the original, 0.728% in WeSpeaker, 0.94% in the ReDimNet2 table, and 1.12% in the 2026 Fair-Gate paper. The w2v-BERT 2.0 authors report 0.12%, while the ReDimNet2 comparison table lists 0.14% for a w2v-BERT 2.0 model trained on VoxBlink2 plus VoxCeleb2, and we could not match those two entries exactly. The VoxBlink2 paper’s stated 43.4% EER reduction does not match its table (0.606% to 0.340% is 43.9%), so we do not quote it. Its abstract also prints minDCF as 0.006%, where the table prints 0.006. The CN-Celeb paper gives 15.52% in its introduction and 16.59% in its table for CN-Celeb.E; we use the table. Older results here (2017 to mid-2023) are flagged as most recent available where no newer equivalent exists. NIST results for the next evaluation cycle would change section 6 first. All EER values here concern other real speakers as impostors, so they say nothing about resistance to synthetic voices.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when NIST publishes numeric results from its next Speaker Recognition Evaluation.