System Detail

wespeaker/w2vbert2

vb2-w2vbert

Submitted by: Public checkpoint

ranked review: approved WeSpeaker W2V-BERT 2.0 + MFA W2V-BERT 2.0

Global position: #6 / 23
Scenario-macro EER: 9.9604
Scenario-macro minDCF: 0.360767
Trial-set coverage: 26 / 26
Scenario top-three: 1

Publicly ranked This result is included in the reviewed public ranking.

Global leaderboard Compare variants Submit another run

Model Provenance

Source and architecture

Displayed name: wespeaker/w2vbert2
Submitted by: Public checkpoint
Metadata: OpenSVBench model record
Source: WeSpeaker
Architecture: W2V-BERT 2.0 + MFA
Model family: W2V-BERT 2.0
Submitted alias: voxceleb_voxblink2_w2v_bert2_lora_adapterMFA_lm
Created at: 2026-05-29T14:24:45.970729+00:00

Training And Links

Training data and references

Training data: VoxCeleb2 dev + VoxBlink2
Training setup: W2V-BERT 2.0 speaker-verification recipe with a 24-layer ~600M Conformer frontend and 80-dimensional normalized filterbanks. The MFA backend uses all 25 hidden states with 128-dimensional Layer Adapter compression and a 256-dimensional embedding. Stage 1 freezes the pretrained frontend and trains LoRA adapters on attention query/value projections (rank 64, alpha 128) with ArcFace (margin 0.2, scale 32), 200–300-frame crops, 0.9/1.1 speed perturbation except for VoxBlink2, and MUSAN/RIR augmentation using AdamW with weight decay 1e-4. Stage 2 jointly fine-tunes the frontend and backend for four epochs with cosine decay from 1e-5 to 5e-6. Stage 3 performs large-margin fine-tuning on VoxCeleb2 only without augmentation, using 500–600-frame crops, margin 0.5, and two epochs. AS-Norm and quality-aware score calibration are applied at evaluation.
Submission mode: full-core
Model source: Open link
Paper: Enhancing Speaker Verification with w2v-BERT 2.0 and Knowledge Distillation guided Structured Pruning

Open Model Source Open Paper

Higher-Ranked Scenarios

Scenario ranks above the global position

Noise/reverb #3 Speaking style #4 Accent/dialect #4

Show 3 supporting trial sets

VoxCeleb1-O: #1 on its trial-set ranking.
VoxPopuli Aging: #1 on its trial-set ranking.
VoxCeleb1-E: #2 on its trial-set ranking.

Lower-Ranked Scenarios

Scenario ranks below the global position

Genre shift #11 Channel/device #9 Distance #7

Show 3 supporting trial sets

FFSVC 2022 Cross-Domain: #15 on its trial-set ranking.
FFSVC 2022 Cross-Channel: #15 on its trial-set ranking.
CN-Celeb Short: #13 on its trial-set ranking.

Scenario Rankings

Full ranking across scenarios

This table shows where the model sits on each scenario, using source-balanced scenario scores and the linked trial sets as evidence.

Scenario	Rank	Lens	Evidence	Score
Noise / Reverb Robustness How reliably the model preserves identity when room acoustics, distractor noise, and microphone placement deviate from an easier in-corpus reference condition.	#3/23 Top 3	Noise and reverberation 0.3200 EER from leader	#3 2.5800	2.5800 0.074040 minDCF
Speaking-Style Robustness Whether the model can preserve identity across emotion-driven change, whispered speech, and noise-induced Lombard speaking style.	#4/23 Competitive	Speaking style shift 0.8728 EER from leader	#5 2.5000 #4 5.8125 #6 0.0559	2.7895 0.157223 minDCF
Accent / Dialect Robustness How reliably the model tracks identity across accent and dialect mismatch.	#4/23 Competitive	Accent and dialect variation 3.6135 EER from leader	#3 24.5937 #5 16.2767	20.4352 0.641166 minDCF
Aging Robustness How stable identity representations remain across age-derived and longitudinal recording gaps.	#5/23 Competitive	Speaker aging and time gaps 0.2929 EER from leader	#5 2.6533 #1 1.5159	2.0846 0.052897 minDCF
Cross-Lingual Robustness How well speaker identity survives enrollment-test language mismatch.	#6/23 Competitive	Language mismatch 2.8100 EER from leader	#6 4.4676	4.4676 0.330255 minDCF
In-The-Wild Robustness How strong the model is on unconstrained celebrity and media speech across official CN-Celeb and VoxCeleb protocols.	#6/23 Competitive	Open-domain media speech 1.4617 EER from leader	#9 9.9690 #1 0.1489 #2 0.3146 #3 0.7298	5.1834 0.162046 minDCF
Overlap Robustness How well speaker identity survives light, mid, and heavy overlap in both meeting and domestic recordings.	#6/23 Competitive	Overlapping speakers 2.0291 EER from leader	#9 7.9267 #4 12.3500	10.1383 0.467346 minDCF
Short-Duration Robustness How much performance holds up when speech evidence is limited by duration.	#6/23 Competitive	Short-duration speech 3.4798 EER from leader	#12 7.0400 #4 6.0250 #13 25.9130 #6 11.0796	12.5144 0.441680 minDCF
Distance Robustness How much performance changes across meeting and domestic distance mismatch.	#7/23 Competitive	Distance mismatch 4.9317 EER from leader	#8 16.0613 #5 6.6900 #3 6.6928 #15 11.3097	10.1885 0.403238 minDCF
Channel / Device Robustness How resilient the model is to device and channel mismatch.	#9/23 Competitive	Device and channel variation 9.2604 EER from leader	#6 15.3867 #15 11.1389	13.2628 0.581970 minDCF
Genre-Shift Robustness Whether performance holds when CN-Celeb enrollment and test speech come from different source genres.	#11/23 Competitive	Source-genre variation 11.6475 EER from leader	#11 25.9200	25.9200 0.656577 minDCF

Variant Compare

Compare W2V-BERT 2.0 variants

Expand this section to compare models in the same model family by architecture, source, training data, training setup, and leaderboard result.

Show

Variant	Submitted by	Source	Training Data	Training Setup	Global Rank	Macro EER	Macro minDCF	Open
W2V-BERT 2.0 + MHFA (VoxCeleb2 + VoxBlink2) W2V-BERT 2.0 + MHFA	pengjy	WeSpeaker	VoxCeleb2 dev + VoxBlink2: 111,301 speakers, 4.83M utterances.	Frontend: facebook/w2v-bert-2.0 (600M). Backend: MHFA with LayerNorm (64 heads, 128-d compression, all 25 hidden states), 256-d embedding. Loss: sub-center ArcFace (K=3, s=32).	#1 ranked	7.5635	0.368244	Open
Dual-LoRA W2V-BERT2-Adapter-MFA (TidyVoice) W2V-BERT 2.0 + Adapter-MFA + Dual-LoRA	shangguanqituan	WeSpeaker	VoxBlink2 and TidyVoiceX training split (3,666 speakers, approximately 262k utterances).	Dual-LoRA adaptation of a frozen W2V-BERT2-Adapter-MFA speaker-verification model pre-trained on VoxBlink2. Independent speaker and language LoRA branches are injected across all 24 transformer layers, using ranks 32 and 16, respectively. A shared language-anchored discriminator with gradient reversal removes linguistic information from the speaker representation. Training uses a three-stage curriculum over three epochs, Sub-center ArcMargin loss (K=3, scale=32, margin increased from 0.2 to 0.3), AdamW, 6-second segments, and MUSAN/RIR augmentation. At inference, only the speaker branch is retained and the speaker LoRA weights can be merged into the backbone.	#4 ranked	9.0606	0.342596	Open
wespeaker/w2vbert2 W2V-BERT 2.0 + MFA Current	Public checkpoint	WeSpeaker	VoxCeleb2 dev + VoxBlink2	W2V-BERT 2.0 speaker-verification recipe with a 24-layer ~600M Conformer frontend and 80-dimensional normalized filterbanks. The MFA backend uses all 25 hidden states with 128-dimensional Layer Adapter compression and a 256-dimensional embedding. Stage 1 freezes the pretrained frontend and trains LoRA adapters on attention query/value projections (rank 64, alpha 128) with ArcFace (margin 0.2, scale 32), 200–300-frame crops, 0.9/1.1 speed perturbation except for VoxBlink2, and MUSAN/RIR augmentation using AdamW with weight decay 1e-4. Stage 2 jointly fine-tunes the frontend and backend for four epochs with cosine decay from 1e-5 to 5e-6. Stage 3 performs large-margin fine-tuning on VoxCeleb2 only without augmentation, using 500–600-frame crops, margin 0.5, and two epochs. AS-Norm and quality-aware score calibration are applied at evaluation.	#6 ranked	9.9604	0.360767	Open

Trial-Set Results

Raw ranking by trial set

Expand this section to inspect the detailed trial-set rankings.

Show

Trial Set	Rank	Scenarios	Trials	Score
TidyVoiceX2-ASV TidyVoiceX2-ASV	#6/23	Cross-lingual	200000	4.4676 0.330255 minDCF
HI-MIA HI-MIA	#12/23	Short-duration	660000	7.0400 0.349597 minDCF
GSC Short Google Speech Commands	#4/23	Short-duration	220000	6.0250 0.388025 minDCF
GLOBE GLOBE	#3/23	Accent/dialect	56848	24.5937 0.512152 minDCF
CN-Celeb CN-Celeb	#9/23	In-the-wild	3484292	9.9690 0.300308 minDCF
CN-Celeb Genre CN-Celeb	#11/23	Genre shift	440000	25.9200 0.656577 minDCF
CN-Celeb Short CN-Celeb	#13/23	Short-duration	546964	25.9130 0.658933 minDCF
VoxCeleb1-O VoxCeleb1	#1/23	In-the-wild	37611	0.1489 0.010954 minDCF
VoxCeleb1-E VoxCeleb1	#2/23	In-the-wild	579818	0.3146 0.018750 minDCF
VoxCeleb1-H VoxCeleb1	#3/23	In-the-wild	550894	0.7298 0.041650 minDCF
VoxCeleb Short VoxCeleb1	#6/23	Short-duration	394724	11.0796 0.370165 minDCF
3D-Speaker Device 3D-Speaker	#6/23	Channel/device	180000	15.3867 0.634773 minDCF
3D-Speaker Distance 3D-Speaker	#8/23	Distance	175163	16.0613 0.619447 minDCF
3D-Speaker Dialect 3D-Speaker	#5/23	Accent/dialect	180000	16.2767 0.770180 minDCF
FFSVC 2022 Cross-Channel FFSVC 2022	#15/23	Channel/device	72000	11.1389 0.529167 minDCF
FFSVC 2022 Cross-Domain FFSVC 2022	#15/23	Distance	66546	11.3097 0.539530 minDCF
Whisper40 Whisper Whisper40	#4/23	Speaking style	17600	5.8125 0.291250 minDCF
Lombard Grid Lombard Lombard Grid	#6/23	Speaking style	29524	0.0559 0.005104 minDCF
VOiCES Noise/Reverb VOiCES	#3/23	Noise/reverb	55000	2.5800 0.074040 minDCF
CHiME-6 Domestic Far-Field CHiME-6	#3/23	Distance	36487	6.6928 0.260295 minDCF
CHiME-6 Overlap CHiME-6	#4/23	Overlap	39600	12.3500 0.603472 minDCF
ESD ESD	#5/23	Speaking style	437408	2.5000 0.175313 minDCF
AliMeeting Near/Far AliMeeting	#5/23	Distance	220000	6.6900 0.193680 minDCF
AliMeeting Overlap AliMeeting	#9/23	Overlap	165000	7.9267 0.331220 minDCF
VoxKnesset VoxKnesset	#5/23	Aging	158312	2.6533 0.081689 minDCF
VoxPopuli Aging VoxPopuli	#1/23	Aging	146575	1.5159 0.024105 minDCF