System Detail

W2V-BERT 2.0 + MHFA (VoxCeleb2 + VoxBlink2)

Wav2Vec2-BERT

Submitted by: pengjy

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

Global position: #4 / 25
Scenario-macro EER: 7.5635
Scenario-macro minDCF: 0.368244
Trial-set coverage: 26 / 26
Scenario top-three: 4

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: W2V-BERT 2.0 + MHFA (VoxCeleb2 + VoxBlink2)
Submitted by: pengjy
Metadata: OpenSVBench model record
Source: WeSpeaker
Architecture: W2V-BERT 2.0 + MHFA
Model family: W2V-BERT 2.0
Submitted alias: BUT/w2v-bert-2.0-mhfa-vox2-vb2
Created at: 2026-10-18T20:56:28.807794+00:00

Training And Links

Training data and references

Training data: VoxCeleb2 dev + VoxBlink2: 111,301 speakers, 4.83M utterances.
Training setup: 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).
Score processing: QMF / quality-based adjustment
Submission mode: full-core
Model source: Open link
Paper: An Attention-Based Backend Allowing Efficient Fine-Tuning of Transformer Models for Speaker Verification

Open Model Source Open Paper

Higher-Ranked Scenarios

Scenario ranks above the global position

Aging #1 Speaking style #1 Overlap #2

Show 3 supporting trial sets

VoxKnesset: #1 on its trial-set ranking.
Whisper40 Whisper: #1 on its trial-set ranking.
CHiME-6 Overlap: #1 on its trial-set ranking.

Lower-Ranked Scenarios

Scenario ranks below the global position

Cross-lingual #7 Channel/device #6 Accent/dialect #5

Show 3 supporting trial sets

GSC Short: #9 on its trial-set ranking.
TidyVoiceX2-ASV: #7 on its trial-set ranking.
VoxCeleb1-O: #7 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
Aging Robustness How stable identity representations remain across age-derived and longitudinal recording gaps.	#1/25 Leader	Speaker aging and time gaps 0.0000 EER from leader	#1 2.0600 #6 1.5235	1.7917 0.060666 minDCF
Speaking-Style Robustness Whether the model can preserve identity across emotion-driven change, whispered speech, and noise-induced Lombard speaking style.	#1/25 Leader	Speaking style shift 0.0000 EER from leader	#5 1.7975 #1 3.9375 #2 0.0149	1.9166 0.153402 minDCF
Overlap Robustness How well speaker identity survives light, mid, and heavy overlap in both meeting and domestic recordings.	#2/25 Top 3	Overlapping speakers 0.4176 EER from leader	#5 5.1047 #1 11.1139	8.1093 0.560809 minDCF
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.	#3/25 Top 3	Noise and reverberation 0.1400 EER from leader	#3 2.4000	2.4000 0.086400 minDCF
Short-Duration Robustness How much performance holds up when speech evidence is limited by duration.	#4/25 Competitive	Short-duration speech 0.7464 EER from leader	#5 3.6067 #9 7.0645 #2 17.5046 #2 9.1879	9.3409 0.524054 minDCF
In-The-Wild Robustness How strong the model is on unconstrained celebrity and media speech across official CN-Celeb and VoxCeleb protocols.	#5/25 Competitive	Open-domain media speech 1.3850 EER from leader	#6 7.0671 #7 0.2658 #6 0.4005 #5 0.7583	3.7710 0.155142 minDCF
Distance Robustness How much performance changes across meeting and domestic distance mismatch.	#5/25 Competitive	Distance mismatch 2.1218 EER from leader	#5 9.6287 #6 5.4400 #2 6.5421 #5 5.6533	6.8160 0.373691 minDCF
Genre-Shift Robustness Whether performance holds when CN-Celeb enrollment and test speech come from different source genres.	#5/25 Competitive	Source-genre variation 5.7605 EER from leader	#5 19.0005	19.0005 0.737320 minDCF
Accent / Dialect Robustness How reliably the model tracks identity across accent and dialect mismatch.	#5/25 Competitive	Accent and dialect variation 3.4229 EER from leader	#1 23.6455 #6 14.7233	19.1844 0.677736 minDCF
Channel / Device Robustness How resilient the model is to device and channel mismatch.	#6/25 Competitive	Device and channel variation 4.2416 EER from leader	#6 10.8460 #5 5.5694	8.2077 0.509507 minDCF
Cross-Lingual Robustness How well speaker identity survives enrollment-test language mismatch.	#7/25 Competitive	Language mismatch 1.0028 EER from leader	#7 2.6605	2.6605 0.211953 minDCF

Variant Compare

Compare W2V-BERT 2.0 variants

Compare the family variants on the overall board, a Scenario, or a Trial Set.

Compare by

Show

Variant	Submitted by	Source	Training Data	Training Setup	Rank	EER	minDCF	Open
w2v-bert2.0 + cc + ecapa-tdnn + astp + sf2 W2V-BERT 2.0 + MFA	JinghanPeng	Contributor-provided	VoxCeleb2 dev + VoxBlink2 + CnCeleb + TidyVoiceX train + 3D-Speaker train + KeSpeech	SphereFace2-C Loss Function + MUSAN/RIRS augmentation	#1	6.0847	0.235010	Open
W2V-BERT 2.0 + Adapter-MFA (cross-device consistency FT + weight soup) W2V-BERT 2.0 + Adapter-MFA	ActionPower	WeSpeaker	Init/anchor: public wespeaker W2V-BERT 2.0 + MFA checkpoint (VoxCeleb2 dev + VoxBlink2). Fine-tuning (~40k speakers): VoxCeleb2 dev, CN-Celeb1 dev + CN-Celeb2, 3D-Speaker train, AliMeeting Train (near + far), TidyVoiceX train, Omnilingual ASR corpus, WaxalNLP, and Korean speech corpora (AI-Hub).	LoRA fine-tuning + ArcFace + cross-device consistency loss (parallel multi-device recordings) + augmentation consistency + model soup with the public wespeaker checkpoint	#2	7.1907	0.288926	Open
W2V-BERT 2.0 + MHFA (VoxCeleb2 + VoxBlink2) W2V-BERT 2.0 + MHFA Current	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).	#4	7.5635	0.368244	Open
Dual-LoRA W2V-BERT2-Adapter-MFA (TidyVoice) W2V-BERT 2.0 + Adapter-MFA + Dual-LoRA	sgqt	WeSpeaker	VoxCeleb2 dev + VoxBlink2 + 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 VoxCeleb2 dev and 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.	#7	9.0606	0.342596	Open
wespeaker/w2vbert2 W2V-BERT 2.0 + MFA	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.	#9	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	#7/25	Cross-lingual	200000	2.6605 0.211953 minDCF
HI-MIA HI-MIA	#5/25	Short-duration	660000	3.6067 0.210263 minDCF
GSC Short Google Speech Commands	#9/25	Short-duration	220000	7.0645 0.517445 minDCF
GLOBE GLOBE	#1/25	Accent/dialect	56848	23.6455 0.513913 minDCF
CN-Celeb CN-Celeb	#6/25	In-the-wild	3484292	7.0671 0.277016 minDCF
CN-Celeb Genre CN-Celeb	#5/25	Genre shift	440000	19.0005 0.737320 minDCF
CN-Celeb Short CN-Celeb	#2/25	Short-duration	546964	17.5046 0.797239 minDCF
VoxCeleb1-O VoxCeleb1	#7/25	In-the-wild	37611	0.2658 0.025633 minDCF
VoxCeleb1-E VoxCeleb1	#6/25	In-the-wild	579818	0.4005 0.026122 minDCF
VoxCeleb1-H VoxCeleb1	#5/25	In-the-wild	550894	0.7583 0.048047 minDCF
VoxCeleb Short VoxCeleb1	#2/25	Short-duration	394724	9.1879 0.571269 minDCF
3D-Speaker Device 3D-Speaker	#6/25	Channel/device	180000	10.8460 0.643653 minDCF
3D-Speaker Distance 3D-Speaker	#5/25	Distance	175163	9.6287 0.589211 minDCF
3D-Speaker Dialect 3D-Speaker	#6/25	Accent/dialect	180000	14.7233 0.841560 minDCF
FFSVC 2022 Cross-Channel FFSVC 2022	#5/25	Channel/device	72000	5.5694 0.375361 minDCF
FFSVC 2022 Cross-Domain FFSVC 2022	#5/25	Distance	66546	5.6533 0.381613 minDCF
Whisper40 Whisper Whisper40	#1/25	Speaking style	17600	3.9375 0.310375 minDCF
Lombard Grid Lombard Lombard Grid	#2/25	Speaking style	29524	0.0149 0.001118 minDCF
VOiCES Noise/Reverb VOiCES	#3/25	Noise/reverb	55000	2.4000 0.086400 minDCF
CHiME-6 Domestic Far-Field CHiME-6	#2/25	Distance	36487	6.5421 0.343925 minDCF
CHiME-6 Overlap CHiME-6	#1/25	Overlap	39600	11.1139 0.755278 minDCF
ESD ESD	#5/25	Speaking style	437408	1.7975 0.148715 minDCF
AliMeeting Near/Far AliMeeting	#6/25	Distance	220000	5.4400 0.180015 minDCF
AliMeeting Overlap AliMeeting	#5/25	Overlap	165000	5.1047 0.366340 minDCF
VoxKnesset VoxKnesset	#1/25	Aging	158312	2.0600 0.091088 minDCF
VoxPopuli Aging VoxPopuli	#6/25	Aging	146575	1.5235 0.030244 minDCF