System Detail

MECT-b2-czur-finetuned

Submitted by: czur

ranked review: approved mect MECT-B2 MECT

Global position: #3 / 25
Scenario-macro EER: 7.2276
Scenario-macro minDCF: 0.280633
Trial-set coverage: 26 / 26
Scenario top-three: 6

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: MECT-b2-czur-finetuned
Submitted by: czur
Metadata: Maintainer override
Source: mect
Architecture: MECT-B2
Model family: MECT
Submitted alias: MECT-b2-czur-finetuned
Created at: 2026-11-02T03:34:20.685023+00:00

Training And Links

Training data and references

Training data: VoxCeleb2 dev + VoxBlink2 + 3D-speaker
Training setup: Follow AntSpeaker: MECT-B2 adopts 80-dim mean-normalized log-Mel filterbanks (16 kHz, 25 ms window, 10 ms hop, 20–7600 Hz). Its 4-stage CNN-Transformer has C=32, M=3, d=64. It employs frame-level 4-expert MoE, weighted dense fusion, attentive stats pooling and 192-dim projection, trained via SphereFace2. It is further fine-tuned on 3D-Speaker to optimize for dialect variations and varying recording distances.
Score processing: None
Submission mode: full-core
Model source: Open link
Paper: MECT: Mixture of Experts with CNN-Transformer Network for Speaker Verification

Open Model Source Open Paper

Higher-Ranked Scenarios

Scenario ranks above the global position

Channel/device #1 Accent/dialect #2 Cross-lingual #3

Show 3 supporting trial sets

HI-MIA: #1 on its trial-set ranking.
3D-Speaker Dialect: #1 on its trial-set ranking.
FFSVC 2022 Cross-Channel: #2 on its trial-set ranking.

Lower-Ranked Scenarios

Scenario ranks below the global position

Aging #13 Noise/reverb #11 Speaking style #8

Show 3 supporting trial sets

VoxCeleb1-E: #18 on its trial-set ranking.
VoxCeleb1-O: #18 on its trial-set ranking.
VoxCeleb1-H: #16 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
Channel / Device Robustness How resilient the model is to device and channel mismatch.	#1/25 Leader	Device and channel variation 0.0000 EER from leader	#2 4.6767 #2 3.2556	3.9661 0.230757 minDCF
Accent / Dialect Robustness How reliably the model tracks identity across accent and dialect mismatch.	#2/25 Top 3	Accent and dialect variation 0.1736 EER from leader	#13 25.1935 #1 6.6767	15.9351 0.447792 minDCF
Cross-Lingual Robustness How well speaker identity survives enrollment-test language mismatch.	#3/25 Top 3	Language mismatch 0.2102 EER from leader	#3 1.8679	1.8679 0.126853 minDCF
In-The-Wild Robustness How strong the model is on unconstrained celebrity and media speech across official CN-Celeb and VoxCeleb protocols.	#3/25 Top 3	Open-domain media speech 1.0559 EER from leader	#4 5.6266 #18 0.9095 #18 1.1318 #16 1.7309	3.4420 0.154836 minDCF
Distance Robustness How much performance changes across meeting and domestic distance mismatch.	#3/25 Top 3	Distance mismatch 0.5786 EER from leader	#2 5.5433 #4 4.2450 #9 7.9891 #2 3.3137	5.2728 0.253762 minDCF
Overlap Robustness How well speaker identity survives light, mid, and heavy overlap in both meeting and domestic recordings.	#3/25 Top 3	Overlapping speakers 0.7054 EER from leader	#2 4.0720 #8 12.7222	8.3971 0.410954 minDCF
Short-Duration Robustness How much performance holds up when speech evidence is limited by duration.	#5/25 Competitive	Short-duration speech 1.3055 EER from leader	#1 2.3383 #7 6.3860 #5 18.2769 #13 12.5989	9.9000 0.379431 minDCF
Genre-Shift Robustness Whether performance holds when CN-Celeb enrollment and test speech come from different source genres.	#7/25 Competitive	Source-genre variation 6.4000 EER from leader	#7 19.6400	19.6400 0.649620 minDCF
Speaking-Style Robustness Whether the model can preserve identity across emotion-driven change, whispered speech, and noise-induced Lombard speaking style.	#8/25 Competitive	Speaking style shift 1.1024 EER from leader	#6 2.1225 #7 6.8750 #8 0.0596	3.0190 0.171062 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.	#11/25 Competitive	Noise and reverberation 2.8220 EER from leader	#11 5.0820	5.0820 0.139280 minDCF
Aging Robustness How stable identity representations remain across age-derived and longitudinal recording gaps.	#13/25 Competitive	Speaker aging and time gaps 1.1894 EER from leader	#13 4.2962 #11 1.6660	2.9811 0.122619 minDCF

Variant Compare

Compare MECT 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
MECT-b2-czur-finetuned MECT-B2 Current	czur	mect	VoxCeleb2 dev + VoxBlink2 + 3D-speaker	Follow AntSpeaker: MECT-B2 adopts 80-dim mean-normalized log-Mel filterbanks (16 kHz, 25 ms window, 10 ms hop, 20–7600 Hz). Its 4-stage CNN-Transformer has C=32, M=3, d=64. It employs frame-level 4-expert MoE, weighted dense fusion, attentive stats pooling and 192-dim projection, trained via SphereFace2. It is further fine-tuned on 3D-Speaker to optimize for dialect variations and varying recording distances.	#3	7.2276	0.280633	Open
AntSpeaker/MECT-B2 MECT-B2	Public checkpoint	AntSpeaker	VoxCeleb2 dev + VoxBlink2	MECT-B2 uses 80-dimensional mean-normalized log-Mel filterbanks at 16 kHz with a 25 ms window, 10 ms shift, and 20–7600 Hz band, followed by a four-stage CNN-Transformer with C=32, M=3, and Transformer width d=64. The model uses frame-level dense mixture-of-experts with four experts, weighted dense fusion, attentive statistics pooling, and a 192-dimensional projection trained with SphereFace2. The VoxCeleb2 + VoxBlink2 setting uses full training followed by large-margin fine-tuning without speed perturbation; the LMF stage uses VoxCeleb2.	#6	8.5599	0.321518	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	#3/25	Cross-lingual	200000	1.8679 0.126853 minDCF
HI-MIA HI-MIA	#1/25	Short-duration	660000	2.3383 0.120418 minDCF
GSC Short Google Speech Commands	#7/25	Short-duration	220000	6.3860 0.353380 minDCF
GLOBE GLOBE	#13/25	Accent/dialect	56848	25.1935 0.495685 minDCF
CN-Celeb CN-Celeb	#4/25	In-the-wild	3484292	5.6266 0.231593 minDCF
CN-Celeb Genre CN-Celeb	#7/25	Genre shift	440000	19.6400 0.649620 minDCF
CN-Celeb Short CN-Celeb	#5/25	Short-duration	546964	18.2769 0.594276 minDCF
VoxCeleb1-O VoxCeleb1	#18/25	In-the-wild	37611	0.9095 0.066421 minDCF
VoxCeleb1-E VoxCeleb1	#18/25	In-the-wild	579818	1.1318 0.070483 minDCF
VoxCeleb1-H VoxCeleb1	#16/25	In-the-wild	550894	1.7309 0.097331 minDCF
VoxCeleb Short VoxCeleb1	#13/25	Short-duration	394724	12.5989 0.449649 minDCF
3D-Speaker Device 3D-Speaker	#2/25	Channel/device	180000	4.6767 0.273987 minDCF
3D-Speaker Distance 3D-Speaker	#2/25	Distance	175163	5.5433 0.321723 minDCF
3D-Speaker Dialect 3D-Speaker	#1/25	Accent/dialect	180000	6.6767 0.399900 minDCF
FFSVC 2022 Cross-Channel FFSVC 2022	#2/25	Channel/device	72000	3.2556 0.187528 minDCF
FFSVC 2022 Cross-Domain FFSVC 2022	#2/25	Distance	66546	3.3137 0.191582 minDCF
Whisper40 Whisper Whisper40	#7/25	Speaking style	17600	6.8750 0.401000 minDCF
Lombard Grid Lombard Lombard Grid	#8/25	Speaking style	29524	0.0596 0.003279 minDCF
VOiCES Noise/Reverb VOiCES	#11/25	Noise/reverb	55000	5.0820 0.139280 minDCF
CHiME-6 Domestic Far-Field CHiME-6	#9/25	Distance	36487	7.9891 0.352608 minDCF
CHiME-6 Overlap CHiME-6	#8/25	Overlap	39600	12.7222 0.594389 minDCF
ESD ESD	#6/25	Speaking style	437408	2.1225 0.108907 minDCF
AliMeeting Near/Far AliMeeting	#4/25	Distance	220000	4.2450 0.149135 minDCF
AliMeeting Overlap AliMeeting	#2/25	Overlap	165000	4.0720 0.227520 minDCF
VoxKnesset VoxKnesset	#13/25	Aging	158312	4.2962 0.199745 minDCF
VoxPopuli Aging VoxPopuli	#11/25	Aging	146575	1.6660 0.045493 minDCF