System Detail

w2v-bert2.0 + cc + ecapa-tdnn + astp + sf2

w2v-bert2.0 + cc + ecapa-tdnn + astp + sf2

Submitted by: JinghanPeng

ranked review: approved Contributor-provided Contributor-provided architecture W2V-BERT 2.0
Global position
#3 / 26
Scenario-macro EER
6.0847
Scenario-macro minDCF
0.235010
Trial-set coverage
26 / 26
Scenario top-three
8
Publicly ranked This result is included in the reviewed public ranking.
Model Provenance

Source and architecture

  • Displayed name: w2v-bert2.0 + cc + ecapa-tdnn + astp + sf2
  • Submitted by: JinghanPeng
  • Metadata: Contributor-provided
  • Source: Contributor-provided
  • Contributor-provided architecture: W2V-BERT 2.0 + MFA
  • Model family: W2V-BERT 2.0
  • Submitted alias: w2v-bert2.0 + cc + ecapa-tdnn + astp + sf2
  • Created at: 2026-11-12T05:33:29.817940+00:00
Training And Links

Training data and references

Higher-Ranked Scenarios

Scenario ranks above the global position

Show 3 supporting trial sets
  • CHiME-6 Overlap: #1 on its trial-set ranking.
  • VoxCeleb1-O: #2 on its trial-set ranking.
  • VoxCeleb1-E: #2 on its trial-set ranking.
Lower-Ranked Scenarios

Scenario ranks below the global position

Show 3 supporting trial sets
  • VoxCeleb Short: #10 on its trial-set ranking.
  • VOiCES Noise/Reverb: #10 on its trial-set ranking.
  • HI-MIA: #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
Overlap Robustness
How well speaker identity survives light, mid, and heavy overlap in both meeting and domestic recordings.
#1/26 Leader
Overlapping speakers 0.0000 EER from leader
7.6917
0.398388 minDCF
In-The-Wild Robustness
How strong the model is on unconstrained celebrity and media speech across official CN-Celeb and VoxCeleb protocols.
#2/26 Top 3
Open-domain media speech 0.3709 EER from leader
2.3860
0.102132 minDCF
Distance Robustness
How much performance changes across meeting and domestic distance mismatch.
#2/26 Top 3
Distance mismatch 0.6761 EER from leader
4.6942
0.223632 minDCF
Cross-Lingual Robustness
How well speaker identity survives enrollment-test language mismatch.
#3/26 Top 3
Language mismatch 0.2044 EER from leader
1.8522
0.125335 minDCF
Speaking-Style Robustness
Whether the model can preserve identity across emotion-driven change, whispered speech, and noise-induced Lombard speaking style.
#3/26 Top 3
Speaking style shift 0.2391 EER from leader
1.9866
0.119754 minDCF
Short-Duration Robustness
How much performance holds up when speech evidence is limited by duration.
#3/26 Top 3
Short-duration speech 2.2797 EER from leader
8.5945
0.315268 minDCF
Genre-Shift Robustness
Whether performance holds when CN-Celeb enrollment and test speech come from different source genres.
#3/26 Top 3
Source-genre variation 4.6067 EER from leader
13.2400
0.461908 minDCF
Accent / Dialect Robustness
How reliably the model tracks identity across accent and dialect mismatch.
#3/26 Top 3
Accent and dialect variation 0.7630 EER from leader
15.7615
0.443685 minDCF
Aging Robustness
How stable identity representations remain across age-derived and longitudinal recording gaps.
#5/26 Competitive
Speaker aging and time gaps 0.2062 EER from leader
2.0580
0.052769 minDCF
Channel / Device Robustness
How resilient the model is to device and channel mismatch.
#5/26 Competitive
Device and channel variation 1.0247 EER from leader
4.0075
0.235878 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.
#10/26 Competitive
Noise and reverberation 2.4000 EER from leader
4.6600
0.106360 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 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
Current
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
#3/26
Cross-lingual
200000
1.8522
0.125335 minDCF
HI-MIA
HI-MIA
#7/26
Short-duration
660000
3.6515
0.192412 minDCF
GSC Short
Google Speech Commands
#4/26
Short-duration
220000
3.6750
0.218370 minDCF
GLOBE
GLOBE
#5/26
Accent/dialect
56848
24.6130
0.505437 minDCF
CN-Celeb
CN-Celeb
#3/26
In-the-wild
3484292
4.4044
0.182453 minDCF
CN-Celeb Genre
CN-Celeb
#3/26
Genre shift
440000
13.2400
0.461908 minDCF
CN-Celeb Short
CN-Celeb
#3/26
Short-duration
546964
15.7791
0.493977 minDCF
VoxCeleb1-O
VoxCeleb1
#2/26
In-the-wild
37611
0.1542
0.009360 minDCF
VoxCeleb1-E
VoxCeleb1
#2/26
In-the-wild
579818
0.3046
0.018716 minDCF
VoxCeleb1-H
VoxCeleb1
#2/26
In-the-wild
550894
0.6443
0.037355 minDCF
VoxCeleb Short
VoxCeleb1
#10/26
Short-duration
394724
11.2724
0.356315 minDCF
3D-Speaker Device
3D-Speaker
#3/26
Channel/device
180000
4.0567
0.231867 minDCF
3D-Speaker Distance
3D-Speaker
#2/26
Distance
175163
4.6378
0.265174 minDCF
3D-Speaker Dialect
3D-Speaker
#3/26
Accent/dialect
180000
6.9100
0.381933 minDCF
FFSVC 2022 Cross-Channel
FFSVC 2022
#5/26
Channel/device
72000
3.9583
0.239889 minDCF
FFSVC 2022 Cross-Domain
FFSVC 2022
#5/26
Distance
66546
4.0192
0.243857 minDCF
Whisper40 Whisper
Whisper40
#3/26
Speaking style
17600
4.1250
0.245563 minDCF
Lombard Grid Lombard
Lombard Grid
#5/26
Speaking style
29524
0.0373
0.001826 minDCF
VOiCES Noise/Reverb
VOiCES
#10/26
Noise/reverb
55000
4.6600
0.106360 minDCF
CHiME-6 Domestic Far-Field
CHiME-6
#5/26
Distance
36487
7.1149
0.257522 minDCF
CHiME-6 Overlap
CHiME-6
#1/26
Overlap
39600
11.3500
0.572056 minDCF
ESD
ESD
#4/26
Speaking style
437408
1.7975
0.111873 minDCF
AliMeeting Near/Far
AliMeeting
#2/26
Distance
220000
3.0050
0.127975 minDCF
AliMeeting Overlap
AliMeeting
#2/26
Overlap
165000
4.0333
0.224720 minDCF
VoxKnesset
VoxKnesset
#6/26
Aging
158312
2.6000
0.080781 minDCF
VoxPopuli Aging
VoxPopuli
#3/26
Aging
146575
1.5159
0.024758 minDCF