System Detail
Agora Voiceprint (ResNet293)
agora_vp
Submitted by: agora_vp
- Global position
- #2 / 23
- Scenario-macro EER
- 7.8931
- Scenario-macro minDCF
- 0.297508
- Trial-set coverage
- 26 / 26
- Scenario top-three
- 5
Publicly ranked
This result is included in the reviewed public ranking.
Model Provenance
Source and architecture
- Displayed name:
Agora Voiceprint (ResNet293) - Submitted by: agora_vp
- Metadata: Contributor-provided
- Source: custom_local_torch
- Contributor-provided architecture: ResNet293 + TSTP (256-d embedding)
- Model family: ResNet
- Submitted alias:
agora_voiceprint - Created at:
2026-10-21T10:53:09.348648+00:00
Training And Links
Training data and references
- Training data: VoxCeleb2 dev + VoxBlink2 + 3D-Speaker train split
- Training setup: AAM-Softmax / ArcMargin Speed perturbation MUSAN noise + RIR reverberation augmentation Non-VAD diff-chunk Knowledge distillation
- Submission mode:
full-core
Higher-Ranked Scenarios
Scenario ranks above the global position
Show 3 supporting trial sets
- HI-MIA: #1 on its trial-set ranking.
- GSC Short: #1 on its trial-set ranking.
- FFSVC 2022 Cross-Channel: #1 on its trial-set ranking.
Lower-Ranked Scenarios
Scenario ranks below the global position
Show 3 supporting trial sets
- VOiCES Noise/Reverb: #13 on its trial-set ranking.
- VoxKnesset: #10 on its trial-set ranking.
- VoxCeleb1-O: #10 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.
|
#1/23
Leader
|
Device and channel variation
0.0000 EER from leader
|
4.0023
0.253548 minDCF
|
|
|
Distance Robustness
How much performance changes across meeting and domestic distance mismatch.
|
#1/23
Leader
|
Distance mismatch
0.0000 EER from leader
|
5.2568
0.269814 minDCF
|
|
|
Short-Duration Robustness
How much performance holds up when speech evidence is limited by duration.
|
#1/23
Leader
|
Short-duration speech
0.0000 EER from leader
|
9.0346
0.315093 minDCF
|
|
|
Accent / Dialect Robustness
How reliably the model tracks identity across accent and dialect mismatch.
|
#1/23
Leader
|
Accent and dialect variation
0.0000 EER from leader
|
16.8217
0.520196 minDCF
|
|
|
Cross-Lingual Robustness
How well speaker identity survives enrollment-test language mismatch.
|
#2/23
Top 3
|
Language mismatch
0.7277 EER from leader
|
2.3854
0.170770 minDCF
|
|
|
Overlap Robustness
How well speaker identity survives light, mid, and heavy overlap in both meeting and domestic recordings.
|
#4/23
Competitive
|
Overlapping speakers
1.2766 EER from leader
|
9.3858
0.412967 minDCF
|
|
|
In-The-Wild Robustness
How strong the model is on unconstrained celebrity and media speech across official CN-Celeb and VoxCeleb protocols.
|
#5/23
Competitive
|
Open-domain media speech
1.1625 EER from leader
|
4.8843
0.165519 minDCF
|
|
|
Genre-Shift Robustness
Whether performance holds when CN-Celeb enrollment and test speech come from different source genres.
|
#5/23
Competitive
|
Source-genre variation
7.7100 EER from leader
|
21.9825
0.674857 minDCF
|
|
|
Speaking-Style Robustness
Whether the model can preserve identity across emotion-driven change, whispered speech, and noise-induced Lombard speaking style.
|
#6/23
Competitive
|
Speaking style shift
2.0661 EER from leader
|
3.9828
0.223228 minDCF
|
|
|
Aging Robustness
How stable identity representations remain across age-derived and longitudinal recording gaps.
|
#9/23
Competitive
|
Speaker aging and time gaps
0.8965 EER from leader
|
2.6882
0.094838 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.
|
#13/23
Needs work
|
Noise and reverberation
4.1400 EER from leader
|
6.4000
0.171760 minDCF
|
Variant Compare
Compare ResNet variants
Expand this section to compare models in the same model family by architecture, source, training data, training setup, and leaderboard result.
Show
Compare ResNet 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 |
|---|---|---|---|---|---|---|---|---|
|
Agora Voiceprint (ResNet293)
ResNet293 + TSTP (256-d embedding)
Current
|
agora_vp | custom_local_torch | VoxCeleb2 dev + VoxBlink2 + 3D-Speaker train split | AAM-Softmax / ArcMargin Speed perturbation MUSAN noise + RIR reverberation augmentation Non-VAD diff-chunk Knowledge distillation |
#2
ranked
|
7.8931 | 0.297508 | Open |
|
wespeaker/res152-voxceleb
ResNet
|
Public checkpoint | WeSpeaker | VoxCeleb2 dev, 5,994 speakers | ResNet152-TSTP-emb256 r-vector, ArcMargin, 150-epoch VoxCeleb2 recipe with speed perturbation and MUSAN/RIRS augmentation. |
#9
ranked
|
11.2275 | 0.442748 | Open |
|
wespeaker/res293-voxceleb
ResNet
|
Public checkpoint | WeSpeaker | VoxCeleb2 dev, 5,994 speakers | ResNet293-TSTP-emb256 r-vector, large-margin fine-tuned; official card reports 28.62M parameters and 28.10G FLOPs. |
#10
ranked
|
11.3130 | 0.439607 | Open |
|
wespeaker/res34-voxceleb
ResNet
|
Public checkpoint | WeSpeaker | VoxCeleb2 dev, 5,994 speakers | ResNet34-TSTP-emb256 r-vector, large-margin fine-tuned; official card reports 6.63M parameters and 4.55G FLOPs. |
#12
ranked
|
12.0516 | 0.477054 | Open |
|
wespeaker/res34-cnceleb
ResNet
|
Public checkpoint | WeSpeaker | CN-Celeb train | ResNet34 r-vector with TSTP pooling and large-margin fine-tuning on the CN-Celeb WeSpeaker recipe. |
#19
ranked
|
14.4910 | 0.607964 | 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
|
#2/23
|
Cross-lingual
|
200000 |
2.3854
0.170770 minDCF
|
|
HI-MIA
HI-MIA
|
#1/23
|
Short-duration
|
660000 |
2.4227
0.143702 minDCF
|
|
GSC Short
Google Speech Commands
|
#1/23
|
Short-duration
|
220000 |
2.5250
0.147370 minDCF
|
|
GLOBE
GLOBE
|
#5/23
|
Accent/dialect
|
56848 |
25.0193
0.505786 minDCF
|
|
CN-Celeb
CN-Celeb
|
#8/23
|
In-the-wild
|
3484292 |
8.9045
0.281688 minDCF
|
|
CN-Celeb Genre
CN-Celeb
|
#5/23
|
Genre shift
|
440000 |
21.9825
0.674857 minDCF
|
|
CN-Celeb Short
CN-Celeb
|
#6/23
|
Short-duration
|
546964 |
20.8157
0.607910 minDCF
|
|
VoxCeleb1-O
VoxCeleb1
|
#10/23
|
In-the-wild
|
37611 |
0.6170
0.031324 minDCF
|
|
VoxCeleb1-E
VoxCeleb1
|
#9/23
|
In-the-wild
|
579818 |
0.6878
0.043217 minDCF
|
|
VoxCeleb1-H
VoxCeleb1
|
#9/23
|
In-the-wild
|
550894 |
1.2872
0.073510 minDCF
|
|
VoxCeleb Short
VoxCeleb1
|
#4/23
|
Short-duration
|
394724 |
10.3751
0.361392 minDCF
|
|
3D-Speaker Device
3D-Speaker
|
#1/23
|
Channel/device
|
180000 |
4.8713
0.311540 minDCF
|
|
3D-Speaker Distance
3D-Speaker
|
#1/23
|
Distance
|
175163 |
6.6049
0.380226 minDCF
|
|
3D-Speaker Dialect
3D-Speaker
|
#1/23
|
Accent/dialect
|
180000 |
8.6240
0.534607 minDCF
|
|
FFSVC 2022 Cross-Channel
FFSVC 2022
|
#1/23
|
Channel/device
|
72000 |
3.1333
0.195556 minDCF
|
|
FFSVC 2022 Cross-Domain
FFSVC 2022
|
#1/23
|
Distance
|
66546 |
3.1904
0.201044 minDCF
|
|
Whisper40 Whisper
Whisper40
|
#8/23
|
Speaking style
|
17600 |
9.3750
0.516687 minDCF
|
|
Lombard Grid Lombard
Lombard Grid
|
#7/23
|
Speaking style
|
29524 |
0.0708
0.002534 minDCF
|
|
VOiCES Noise/Reverb
VOiCES
|
#13/23
|
Noise/reverb
|
55000 |
6.4000
0.171760 minDCF
|
|
CHiME-6 Domestic Far-Field
CHiME-6
|
#6/23
|
Distance
|
36487 |
7.9289
0.329545 minDCF
|
|
CHiME-6 Overlap
CHiME-6
|
#8/23
|
Overlap
|
39600 |
14.1250
0.591000 minDCF
|
|
ESD
ESD
|
#6/23
|
Speaking style
|
437408 |
2.5025
0.150464 minDCF
|
|
AliMeeting Near/Far
AliMeeting
|
#1/23
|
Distance
|
220000 |
3.3030
0.168440 minDCF
|
|
AliMeeting Overlap
AliMeeting
|
#1/23
|
Overlap
|
165000 |
4.6467
0.234933 minDCF
|
|
VoxKnesset
VoxKnesset
|
#10/23
|
Aging
|
158312 |
3.7554
0.140025 minDCF
|
|
VoxPopuli Aging
VoxPopuli
|
#8/23
|
Aging
|
146575 |
1.6210
0.049651 minDCF
|
Submission Notes
Packaging notes and warnings
Expand this section for metadata, warnings, and missing trial sets.
Show
Packaging notes and warnings
Expand this section for metadata, warnings, and missing trial sets.
Packaging Notes
Submission metadata
- Submission mode:
full-core - Listed on public ranking:
True - Visibility:
public - Review status:
approved - Created at:
2026-10-21T10:53:09.348648+00:00
Warnings
Evaluation notes
- Scores were extracted by a custom local PyTorch adapter with full-utterance FP32 inference; no kit model backend was rerun during packaging.