System Detail

w2vbert2.0+ MHFA

w2vbert2.0+ MHFA

Submitted by: pengjy

ranked review: approved WeSpeaker W2V-BERT 2.0
Global position
#2 / 26
Scenario-macro EER
5.3542
Scenario-macro minDCF
0.228132
Trial-set coverage
26 / 26
Scenario top-three
9
Publicly ranked This result is included in the reviewed public ranking.
Model Provenance

Source and architecture

  • Displayed name: w2vbert2.0+ MHFA
  • Submitted by: pengjy
  • Metadata: Contributor-provided
  • Source: WeSpeaker
  • Model family: W2V-BERT 2.0
  • Submitted alias: BUT/w2v-bert-2.0-mhfa-vox2-vb2
  • Created at: 2026-11-18T08:20:41.425065+00:00
Training And Links

Training data and references

Higher-Ranked Scenarios

Scenario ranks above the global position

Show 3 supporting trial sets
  • AliMeeting Near/Far: #1 on its trial-set ranking.
  • AliMeeting Overlap: #1 on its trial-set ranking.
  • CN-Celeb Genre: #1 on its trial-set ranking.
Lower-Ranked Scenarios

Scenario ranks below the global position

Show 3 supporting trial sets
  • CHiME-6 Overlap: #9 on its trial-set ranking.
  • VoxKnesset: #9 on its trial-set ranking.
  • VoxCeleb1-H: #9 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
Genre-Shift Robustness
Whether performance holds when CN-Celeb enrollment and test speech come from different source genres.
#1/26 Leader
Source-genre variation 0.0000 EER from leader
8.6333
0.346320 minDCF
Speaking-Style Robustness
Whether the model can preserve identity across emotion-driven change, whispered speech, and noise-induced Lombard speaking style.
#2/26 Top 3
Speaking style shift 0.1785 EER from leader
1.9259
0.116187 minDCF
Channel / Device Robustness
How resilient the model is to device and channel mismatch.
#2/26 Top 3
Device and channel variation 0.7272 EER from leader
3.7100
0.243606 minDCF
Short-Duration Robustness
How much performance holds up when speech evidence is limited by duration.
#2/26 Top 3
Short-duration speech 0.0673 EER from leader
6.3821
0.279620 minDCF
Accent / Dialect Robustness
How reliably the model tracks identity across accent and dialect mismatch.
#2/26 Top 3
Accent and dialect variation 0.7535 EER from leader
15.7520
0.471486 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.
#3/26 Top 3
Noise and reverberation 0.1400 EER from leader
2.4000
0.060920 minDCF
In-The-Wild Robustness
How strong the model is on unconstrained celebrity and media speech across official CN-Celeb and VoxCeleb protocols.
#3/26 Top 3
Open-domain media speech 0.6528 EER from leader
2.6679
0.126686 minDCF
Distance Robustness
How much performance changes across meeting and domestic distance mismatch.
#3/26 Top 3
Distance mismatch 0.7037 EER from leader
4.7219
0.244304 minDCF
Overlap Robustness
How well speaker identity survives light, mid, and heavy overlap in both meeting and domestic recordings.
#3/26 Top 3
Overlapping speakers 0.7278 EER from leader
8.4194
0.403614 minDCF
Cross-Lingual Robustness
How well speaker identity survives enrollment-test language mismatch.
#5/26 Competitive
Language mismatch 0.2997 EER from leader
1.9475
0.148847 minDCF
Aging Robustness
How stable identity representations remain across age-derived and longitudinal recording gaps.
#9/26 Competitive
Speaker aging and time gaps 0.4838 EER from leader
2.3355
0.067864 minDCF
Variant Compare

Compare W2V-BERT 2.0 variants

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

Show
Variant Submitted by Source Training Data Training Setup Rank EER minDCF Open
w2vbert2.0+ MHFA abc WeSpeaker DataSet Sampling Share (SpkCount, Hours, NaturalShare) VoxBlink2 25.74% (111,284 spk, 16,122 h, 45.03%) Common Voice 13.07% (67,317 spk, 5,798 h, 22.87%) CN-Celeb 9.86% (2,793 spk, 1,329 h, 2.88%) VoxCeleb2 8.52% (5,994 spk, 2,506 h, 4.96%) TidyVoiceX 8.18% (3,666 spk, 356 h, 1.19%) KeSpeech 7.53% (27,158 spk, 1,076 h, 4.39%) PrivateData 6.03% (19,236 spk, 1,260 h, 2.11%), VOiCES 5.35% (387 spk, 2,947 h, 3.12%), 3D-Speaker 4.51% (10,000 spk, 1,107 h, 2.63%), 3D-Speaker-3A 4.51% (10,000 spk, 3,339 h, 7.90%). toolkit: wespeaker projection: ArcMarginProduct_intertopk_subcenter_uncertianty DataAug: noise (musan+DNS5_noise) + reverberation(DNS5_RIR+RIRS) #1 5.3492 0.267001 Open
w2vbert2.0+ MHFA
Current
pengjy WeSpeaker VoxCeleb2 dev + VoxBlink2 + CnCeleb + TidyVoiceX train + 3D-Speaker train + KeSpeech None #2 5.3542 0.228132 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 #3 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 #4 7.1907 0.288926 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. #8 9.0606 0.342596 Open
W2V-BERT 2.0 + LoRA_Adapter_MFA (VoxCeleb2 dev + VoxBlink2)
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. #10 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
#5/26
Cross-lingual
200000
1.9475
0.148847 minDCF
HI-MIA
HI-MIA
#5/26
Short-duration
660000
3.2567
0.198993 minDCF
GSC Short
Google Speech Commands
#2/26
Short-duration
220000
2.5021
0.148725 minDCF
GLOBE
GLOBE
#2/26
Accent/dialect
56848
24.1293
0.494060 minDCF
CN-Celeb
CN-Celeb
#5/26
In-the-wild
3484292
4.7560
0.216688 minDCF
CN-Celeb Genre
CN-Celeb
#1/26
Genre shift
440000
8.6333
0.346320 minDCF
CN-Celeb Short
CN-Celeb
#1/26
Short-duration
546964
10.0394
0.389685 minDCF
VoxCeleb1-O
VoxCeleb1
#9/26
In-the-wild
37611
0.3137
0.029514 minDCF
VoxCeleb1-E
VoxCeleb1
#9/26
In-the-wild
579818
0.4767
0.027091 minDCF
VoxCeleb1-H
VoxCeleb1
#9/26
In-the-wild
550894
0.9491
0.053448 minDCF
VoxCeleb Short
VoxCeleb1
#4/26
Short-duration
394724
9.7304
0.381078 minDCF
3D-Speaker Device
3D-Speaker
#2/26
Channel/device
180000
3.7200
0.245240 minDCF
3D-Speaker Distance
3D-Speaker
#3/26
Distance
175163
5.0228
0.307259 minDCF
3D-Speaker Dialect
3D-Speaker
#4/26
Accent/dialect
180000
7.3748
0.448913 minDCF
FFSVC 2022 Cross-Channel
FFSVC 2022
#4/26
Channel/device
72000
3.7000
0.241972 minDCF
FFSVC 2022 Cross-Domain
FFSVC 2022
#4/26
Distance
66546
3.7278
0.241743 minDCF
Whisper40 Whisper
Whisper40
#2/26
Speaking style
17600
3.8750
0.199313 minDCF
Lombard Grid Lombard
Lombard Grid
#6/26
Speaking style
29524
0.0373
0.002235 minDCF
VOiCES Noise/Reverb
VOiCES
#3/26
Noise/reverb
55000
2.4000
0.060920 minDCF
CHiME-6 Domestic Far-Field
CHiME-6
#7/26
Distance
36487
7.6877
0.298553 minDCF
CHiME-6 Overlap
CHiME-6
#9/26
Overlap
39600
13.1389
0.594028 minDCF
ESD
ESD
#6/26
Speaking style
437408
1.8656
0.147013 minDCF
AliMeeting Near/Far
AliMeeting
#1/26
Distance
220000
2.4491
0.129660 minDCF
AliMeeting Overlap
AliMeeting
#1/26
Overlap
165000
3.7000
0.213200 minDCF
VoxKnesset
VoxKnesset
#9/26
Aging
158312
3.1400
0.104456 minDCF
VoxPopuli Aging
VoxPopuli
#7/26
Aging
146575
1.5310
0.031272 minDCF