System Detail
palabraai/redimnet2-b6
redimnet2
Submitted by: Public checkpoint
- Global position
- #7 / 23
- Scenario-macro EER
- 10.5277
- Scenario-macro minDCF
- 0.419432
- Trial-set coverage
- 26 / 26
- Scenario top-three
- 0
Publicly ranked
This result is included in the reviewed public ranking.
Model Provenance
Source and architecture
- Displayed name:
palabraai/redimnet2-b6 - Submitted by: Public checkpoint
- Metadata: OpenSVBench model record
- Source: PalabraAI
- Architecture: ReDimNet2-B6
- Model family: ReDimNet2
- Submitted alias:
redimnet2_b6_lm - Created at:
2026-05-29T14:38:17.385092+00:00
Training And Links
Training data and references
- Training data: VoxCeleb2 dev
- Training setup: ReDimNet2-B6 speaker-verification recipe: 80-dimensional mean-normalized log-Mel filterbanks (25 ms window, 10 ms shift), a time-pooled dimension-reshaping backbone with 192-dimensional embeddings, 2-second random crops, MUSAN/RIR augmentation, and 0.9/1.1 speed perturbation during SphereFace2-C training with margin warm-up. Large-margin fine-tuning uses 6-second segments without speed perturbation, fixed margin 0.3, and five epochs with exponential learning-rate decay. The public release reports 12.3M parameters and 13.05 GMACs.
- Submission mode:
full-core - Model source: Open link
- Paper: ReDimNet2: Scaling Speaker Verification via Time-Pooled Dimension Reshaping
Higher-Ranked Scenarios
Scenario ranks above the global position
Show 3 supporting trial sets
- VoxPopuli Aging: #5 on its trial-set ranking.
- VoxCeleb Short: #5 on its trial-set ranking.
- VoxCeleb1-O: #6 on its trial-set ranking.
Lower-Ranked Scenarios
Scenario ranks below the global position
Show 3 supporting trial sets
- AliMeeting Near/Far: #15 on its trial-set ranking.
- CN-Celeb: #14 on its trial-set ranking.
- GSC Short: #13 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.
|
#7/23
Competitive
|
Speaker aging and time gaps
0.6613 EER from leader
|
2.4530
0.088149 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.
|
#7/23
Competitive
|
Noise and reverberation
0.8600 EER from leader
|
3.1200
0.114220 minDCF
|
|
|
Overlap Robustness
How well speaker identity survives light, mid, and heavy overlap in both meeting and domestic recordings.
|
#7/23
Competitive
|
Overlapping speakers
2.5888 EER from leader
|
10.6981
0.524446 minDCF
|
|
|
Accent / Dialect Robustness
How reliably the model tracks identity across accent and dialect mismatch.
|
#7/23
Competitive
|
Accent and dialect variation
4.3113 EER from leader
|
21.1330
0.660821 minDCF
|
|
|
Distance Robustness
How much performance changes across meeting and domestic distance mismatch.
|
#8/23
Competitive
|
Distance mismatch
5.8231 EER from leader
|
11.0799
0.472944 minDCF
|
|
|
Channel / Device Robustness
How resilient the model is to device and channel mismatch.
|
#8/23
Competitive
|
Device and channel variation
8.9729 EER from leader
|
12.9753
0.641859 minDCF
|
|
|
Short-Duration Robustness
How much performance holds up when speech evidence is limited by duration.
|
#8/23
Competitive
|
Short-duration speech
4.2135 EER from leader
|
13.2481
0.531421 minDCF
|
|
|
Cross-Lingual Robustness
How well speaker identity survives enrollment-test language mismatch.
|
#9/23
Competitive
|
Language mismatch
3.1131 EER from leader
|
4.7707
0.339730 minDCF
|
|
|
Speaking-Style Robustness
Whether the model can preserve identity across emotion-driven change, whispered speech, and noise-induced Lombard speaking style.
|
#10/23
Competitive
|
Speaking style shift
2.5336 EER from leader
|
4.4502
0.291976 minDCF
|
|
|
In-The-Wild Robustness
How strong the model is on unconstrained celebrity and media speech across official CN-Celeb and VoxCeleb protocols.
|
#10/23
Competitive
|
Open-domain media speech
2.6998 EER from leader
|
6.4216
0.208137 minDCF
|
|
|
Genre-Shift Robustness
Whether performance holds when CN-Celeb enrollment and test speech come from different source genres.
|
#10/23
Competitive
|
Source-genre variation
11.1825 EER from leader
|
25.4550
0.740052 minDCF
|
Variant Compare
Compare ReDimNet2 variants
Expand this section to compare models in the same model family by architecture, source, training data, training setup, and leaderboard result.
Show
Compare ReDimNet2 variants
Expand this section to compare models in the same model family by architecture, source, training data, training setup, and leaderboard result.
| Variant | Submitted by | Source | Training Data | Training Setup | Global Rank | Macro EER | Macro minDCF | Open |
|---|---|---|---|---|---|---|---|---|
|
palabraai/redimnet2-b6-vb2+vox2
ReDimNet2-B6
|
Public checkpoint | PalabraAI | VoxCeleb2 dev + VoxBlink2 | ReDimNet2-B6 speaker-verification recipe: 80-dimensional mean-normalized log-Mel filterbanks (25 ms window, 10 ms shift), a time-pooled dimension-reshaping backbone with 192-dimensional embeddings, 2-second random crops, MUSAN/RIR augmentation, and 0.9/1.1 speed perturbation during SphereFace2-C training with margin warm-up. Large-margin fine-tuning uses 6-second segments without speed perturbation, fixed margin 0.3, and five epochs with exponential learning-rate decay. This checkpoint is the extended VoxCeleb2 + VoxBlink2 release; the public release reports 12.3M parameters and 13.05 GMACs. |
#5
ranked
|
9.1442 | 0.324160 | Open |
|
palabraai/redimnet2-b6
ReDimNet2-B6
Current
|
Public checkpoint | PalabraAI | VoxCeleb2 dev | ReDimNet2-B6 speaker-verification recipe: 80-dimensional mean-normalized log-Mel filterbanks (25 ms window, 10 ms shift), a time-pooled dimension-reshaping backbone with 192-dimensional embeddings, 2-second random crops, MUSAN/RIR augmentation, and 0.9/1.1 speed perturbation during SphereFace2-C training with margin warm-up. Large-margin fine-tuning uses 6-second segments without speed perturbation, fixed margin 0.3, and five epochs with exponential learning-rate decay. The public release reports 12.3M parameters and 13.05 GMACs. |
#7
ranked
|
10.5277 | 0.419432 | Open |
Trial-Set Results
Raw ranking by trial set
Expand this section to inspect the detailed trial-set rankings.
Show
Raw ranking by trial set
Expand this section to inspect the detailed trial-set rankings.
| Trial Set | Rank | Scenarios | Trials | Score |
|---|---|---|---|---|
|
TidyVoiceX2-ASV
TidyVoiceX2-ASV
|
#9/23
|
Cross-lingual
|
200000 |
4.7707
0.339730 minDCF
|
|
HI-MIA
HI-MIA
|
#6/23
|
Short-duration
|
660000 |
5.7113
0.276288 minDCF
|
|
GSC Short
Google Speech Commands
|
#13/23
|
Short-duration
|
220000 |
12.3950
0.640440 minDCF
|
|
GLOBE
GLOBE
|
#6/23
|
Accent/dialect
|
56848 |
25.0193
0.533495 minDCF
|
|
CN-Celeb
CN-Celeb
|
#14/23
|
In-the-wild
|
3484292 |
12.2444
0.378920 minDCF
|
|
CN-Celeb Genre
CN-Celeb
|
#10/23
|
Genre shift
|
440000 |
25.4550
0.740052 minDCF
|
|
CN-Celeb Short
CN-Celeb
|
#8/23
|
Short-duration
|
546964 |
24.2213
0.769876 minDCF
|
|
VoxCeleb1-O
VoxCeleb1
|
#6/23
|
In-the-wild
|
37611 |
0.2871
0.019358 minDCF
|
|
VoxCeleb1-E
VoxCeleb1
|
#7/23
|
In-the-wild
|
579818 |
0.5198
0.031606 minDCF
|
|
VoxCeleb1-H
VoxCeleb1
|
#8/23
|
In-the-wild
|
550894 |
0.9892
0.061095 minDCF
|
|
VoxCeleb Short
VoxCeleb1
|
#5/23
|
Short-duration
|
394724 |
10.6649
0.439079 minDCF
|
|
3D-Speaker Device
3D-Speaker
|
#8/23
|
Channel/device
|
180000 |
17.2700
0.778607 minDCF
|
|
3D-Speaker Distance
3D-Speaker
|
#10/23
|
Distance
|
175163 |
17.6447
0.753012 minDCF
|
|
3D-Speaker Dialect
3D-Speaker
|
#8/23
|
Accent/dialect
|
180000 |
17.2467
0.788147 minDCF
|
|
FFSVC 2022 Cross-Channel
FFSVC 2022
|
#7/23
|
Channel/device
|
72000 |
8.6806
0.505111 minDCF
|
|
FFSVC 2022 Cross-Domain
FFSVC 2022
|
#8/23
|
Distance
|
66546 |
8.8886
0.517318 minDCF
|
|
Whisper40 Whisper
Whisper40
|
#12/23
|
Speaking style
|
17600 |
10.1875
0.646312 minDCF
|
|
Lombard Grid Lombard
Lombard Grid
|
#9/23
|
Speaking style
|
29524 |
0.0782
0.007675 minDCF
|
|
VOiCES Noise/Reverb
VOiCES
|
#7/23
|
Noise/reverb
|
55000 |
3.1200
0.114220 minDCF
|
|
CHiME-6 Domestic Far-Field
CHiME-6
|
#8/23
|
Distance
|
36487 |
8.4715
0.379560 minDCF
|
|
CHiME-6 Overlap
CHiME-6
|
#7/23
|
Overlap
|
39600 |
13.9722
0.658139 minDCF
|
|
ESD
ESD
|
#9/23
|
Speaking style
|
437408 |
3.0850
0.221939 minDCF
|
|
AliMeeting Near/Far
AliMeeting
|
#15/23
|
Distance
|
220000 |
9.3150
0.241885 minDCF
|
|
AliMeeting Overlap
AliMeeting
|
#8/23
|
Overlap
|
165000 |
7.4240
0.390753 minDCF
|
|
VoxKnesset
VoxKnesset
|
#7/23
|
Aging
|
158312 |
3.3600
0.138451 minDCF
|
|
VoxPopuli Aging
VoxPopuli
|
#5/23
|
Aging
|
146575 |
1.5460
0.037846 minDCF
|