System Detail

Dual-LoRA W2V-BERT2-Adapter-MFA (TidyVoice)

duallora-w2vbert2-tidyvoice-full-core

Submitted by: shangguanqituan

ranked review: approved WeSpeaker Contributor-provided architecture W2V-BERT 2.0

Global position: #4 / 23
Scenario-macro EER: 9.0606
Scenario-macro minDCF: 0.342596
Trial-set coverage: 26 / 26
Scenario top-three: 2

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: Dual-LoRA W2V-BERT2-Adapter-MFA (TidyVoice)
Submitted by: shangguanqituan
Metadata: Contributor-provided
Source: WeSpeaker
Contributor-provided architecture: W2V-BERT 2.0 + Adapter-MFA + Dual-LoRA
Model family: W2V-BERT 2.0
Submitted alias: shangguanqituan/duallora-w2vbert2-tidyvoice
Created at: 2026-10-20T17:21:49.372155+00:00

Training And Links

Training data and references

Training data: VoxBlink2 and TidyVoiceX training split (3,666 speakers, approximately 262k utterances).
Training setup: 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.
Submission mode: full-core
Model source: Open link
Paper: Dual-LoRA: Parameter-Efficient Adversarial Disentanglement for Cross-Lingual Speaker Verification

Open Model Source Open Paper

Higher-Ranked Scenarios

Scenario ranks above the global position

Cross-lingual #1 Speaking style #2 In-the-wild #4

Show 3 supporting trial sets

TidyVoiceX2-ASV: #1 on its trial-set ranking.
ESD: #2 on its trial-set ranking.
GSC Short: #2 on its trial-set ranking.

Lower-Ranked Scenarios

Scenario ranks below the global position

Genre shift #6 Accent/dialect #6 Channel/device #6

Show 3 supporting trial sets

HI-MIA: #20 on its trial-set ranking.
FFSVC 2022 Cross-Domain: #9 on its trial-set ranking.
FFSVC 2022 Cross-Channel: #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
Cross-Lingual Robustness How well speaker identity survives enrollment-test language mismatch.	#1/23 Leader	Language mismatch 0.0000 EER from leader	#1 1.6577	1.6577 0.102286 minDCF
Speaking-Style Robustness Whether the model can preserve identity across emotion-driven change, whispered speech, and noise-induced Lombard speaking style.	#2/23 Top 3	Speaking style shift 0.4333 EER from leader	#2 1.7750 #3 5.2562 #3 0.0186	2.3500 0.152707 minDCF
In-The-Wild Robustness How strong the model is on unconstrained celebrity and media speech across official CN-Celeb and VoxCeleb protocols.	#4/23 Competitive	Open-domain media speech 0.5363 EER from leader	#7 7.8964 #7 0.3776 #8 0.5436 #6 0.9380	4.2581 0.150654 minDCF
Short-Duration Robustness How much performance holds up when speech evidence is limited by duration.	#4/23 Competitive	Short-duration speech 2.6821 EER from leader	#20 10.5467 #2 4.2715 #5 20.2719 #8 11.7768	11.7167 0.436659 minDCF
Overlap Robustness How well speaker identity survives light, mid, and heavy overlap in both meeting and domestic recordings.	#5/23 Competitive	Overlapping speakers 1.7845 EER from leader	#7 7.3153 #5 12.4722	9.8938 0.450429 minDCF
Aging Robustness How stable identity representations remain across age-derived and longitudinal recording gaps.	#6/23 Competitive	Speaker aging and time gaps 0.3746 EER from leader	#6 2.7267 #7 1.6060	2.1663 0.065326 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.	#6/23 Competitive	Noise and reverberation 0.8200 EER from leader	#6 3.0800	3.0800 0.084760 minDCF
Distance Robustness How much performance changes across meeting and domestic distance mismatch.	#6/23 Competitive	Distance mismatch 4.5008 EER from leader	#6 14.9704 #6 7.2850 #5 7.7781 #9 8.9969	9.7576 0.432930 minDCF
Channel / Device Robustness How resilient the model is to device and channel mismatch.	#6/23 Competitive	Device and channel variation 7.8256 EER from leader	#5 14.8253 #9 8.8306	11.8279 0.598239 minDCF
Accent / Dialect Robustness How reliably the model tracks identity across accent and dialect mismatch.	#6/23 Competitive	Accent and dialect variation 3.9367 EER from leader	#2 24.3034 #7 17.2133	20.7584 0.640224 minDCF
Genre-Shift Robustness Whether performance holds when CN-Celeb enrollment and test speech come from different source genres.	#6/23 Competitive	Source-genre variation 7.9275 EER from leader	#6 22.2000	22.2000 0.654345 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 Current	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	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	#1/23	Cross-lingual	200000	1.6577 0.102286 minDCF
HI-MIA HI-MIA	#20/23	Short-duration	660000	10.5467 0.426292 minDCF
GSC Short Google Speech Commands	#2/23	Short-duration	220000	4.2715 0.285070 minDCF
GLOBE GLOBE	#2/23	Accent/dialect	56848	24.3034 0.499381 minDCF
CN-Celeb CN-Celeb	#7/23	In-the-wild	3484292	7.8964 0.263702 minDCF
CN-Celeb Genre CN-Celeb	#6/23	Genre shift	440000	22.2000 0.654345 minDCF
CN-Celeb Short CN-Celeb	#5/23	Short-duration	546964	20.2719 0.627675 minDCF
VoxCeleb1-O VoxCeleb1	#7/23	In-the-wild	37611	0.3776 0.026057 minDCF
VoxCeleb1-E VoxCeleb1	#8/23	In-the-wild	579818	0.5436 0.031696 minDCF
VoxCeleb1-H VoxCeleb1	#6/23	In-the-wild	550894	0.9380 0.055064 minDCF
VoxCeleb Short VoxCeleb1	#8/23	Short-duration	394724	11.7768 0.407599 minDCF
3D-Speaker Device 3D-Speaker	#5/23	Channel/device	180000	14.8253 0.714173 minDCF
3D-Speaker Distance 3D-Speaker	#6/23	Distance	175163	14.9704 0.692739 minDCF
3D-Speaker Dialect 3D-Speaker	#7/23	Accent/dialect	180000	17.2133 0.781067 minDCF
FFSVC 2022 Cross-Channel FFSVC 2022	#9/23	Channel/device	72000	8.8306 0.482306 minDCF
FFSVC 2022 Cross-Domain FFSVC 2022	#9/23	Distance	66546	8.9969 0.496096 minDCF
Whisper40 Whisper Whisper40	#3/23	Speaking style	17600	5.2562 0.320875 minDCF
Lombard Grid Lombard Lombard Grid	#3/23	Speaking style	29524	0.0186 0.001863 minDCF
VOiCES Noise/Reverb VOiCES	#6/23	Noise/reverb	55000	3.0800 0.084760 minDCF
CHiME-6 Domestic Far-Field CHiME-6	#5/23	Distance	36487	7.7781 0.310431 minDCF
CHiME-6 Overlap CHiME-6	#5/23	Overlap	39600	12.4722 0.582611 minDCF
ESD ESD	#2/23	Speaking style	437408	1.7750 0.135385 minDCF
AliMeeting Near/Far AliMeeting	#6/23	Distance	220000	7.2850 0.232455 minDCF
AliMeeting Overlap AliMeeting	#7/23	Overlap	165000	7.3153 0.318247 minDCF
VoxKnesset VoxKnesset	#6/23	Aging	158312	2.7267 0.100356 minDCF
VoxPopuli Aging VoxPopuli	#7/23	Aging	146575	1.6060 0.030296 minDCF