VoxBlink2: Large-Scale Audio-Visual Corpus

• VoxBlink2 is a large-scale audio-visual speaker recognition corpus that significantly expands scale with 111K speakers, 2M videos, 9.9M utterances, and 16,672 hours compared to its predecessor.
• Its multi-stage collection pipeline combines multilingual keyword retrieval, rigorous face and audio processing, and refined diarization to boost label accuracy from 72% to 92% in manual assessments.
• The dataset supports robust training for speaker verification and introduces an open-set identification benchmark, driving substantial EER reductions and state-of-the-art performance in NIST SRE24 systems.

Search arXiv for VoxBlink2, VoxBlink, and NIST SRE24 usages to ground the article in papers. VoxBlink2 is a large-scale audio-visual speaker recognition corpus introduced as the successor to VoxBlink and accompanied by an Open-Set Speaker Identification (OSSI) benchmark. It was designed to push speaker recognition toward data scales more typical of face recognition, and to support both conventional speaker verification and open-set identification with thousands of enrolled and unknown speakers. The dataset contains 111,284 speakers, 2,097,062 videos, 9,904,382 utterances, and 16,672 hours, all sourced from YouTube with paired audio and video (Lin et al., 2024).

1. Dataset definition and scale

VoxBlink2 is explicitly presented as a large audio-visual speaker recognition dataset built from “in-the-wild” YouTube material. Each example is an utterance paired with the corresponding video clip, and the collection targets users who use their own photos as avatars. Relative to earlier public corpora, the dataset is intended to increase speaker count, temporal spread, and scenario diversity rather than merely add more utterances (Lin et al., 2024).

Its scale is the main defining property. VoxBlink2 expands the original VoxBlink from 38,065 speakers, 372,084 videos, 1,455,190 utterances, and 2,135 hours to 111,284 speakers, 2,097,062 videos, 9,904,382 utterances, and 16,672 hours. It also exceeds VoxCeleb2 in speaker count, videos, utterances, and total hours (Lin et al., 2024).

Dataset	Speakers	Utterances
VoxCeleb2	5,994	1,092,009
VoxBlink	38,065	1,455,190
VoxBlink2	111,284	9,904,382

The paper also reports several summary statistics that are relevant for corpus characterization. VoxBlink2 has an average of 89 utterances per speaker, an average utterance duration of 6.0 s, and an average time span per speaker of 786 days. The predecessor VoxBlink had an average span of 440 days. This suggests that VoxBlink2 was designed not only for scale but also for broader longitudinal variability (Lin et al., 2024).

A recurrent misconception is that VoxBlink2 should be interpreted primarily as a blink-specific dataset because of its name. The published description instead defines it as a speaker recognition corpus and benchmark. Downstream speaker-recognition papers treat it as a large-scale supervision source for audio and visual identity learning rather than as a corpus built around explicit blink dynamics (Lin et al., 2024).

2. Collection pipeline and annotation refinement

The collection pipeline is multi-stage and explicitly audio-visual. The initial candidate retrieval uses a long keyword list spanning 18 languages and collects over 6 million 1-minute videos from YouTube users with photo avatars. This broad multilingual retrieval is one of the main differences from the original VoxBlink, which relied more heavily on short segments (Lin et al., 2024).

Frame extraction is performed at 25 fps. Face detection uses MobileNet, and single-person video tracks are built using a minimum Intersection over Union (IOU) threshold between consecutive units so that each facial track contains only one person. Face identity assignment then uses a pre-trained ArcFace classifier. The authors explicitly state that moving from a verification-style stage to a stronger identification stage improved data purity (Lin et al., 2024).

The audio side of curation is not treated independently. VoxBlink2 applies an audio-visual speaker diarization model together with an overlap speech detection model to partition active speech segments and remove silent and overlapped speech segments. This is a substantive design choice: the corpus is not merely face-filtered video, but a collection of active speaking segments with overlap suppression (Lin et al., 2024).

A second branch further improves identity purity by training a face classifier on mined data. That branch uses ResNet-IRSE50, compares face embeddings against the user’s avatar photo, samples up to 10 faces per speaker using cosine-similarity-based weighted sampling, and appends ArcFace to the encoder for classifier training on about 200K candidate face data. The reported outcome is a manual accuracy increase from 72% for VoxBlink to 92% for VoxBlink2, based on manual assessment of 50 randomly sampled speakers (Lin et al., 2024).

This quality estimate should be interpreted carefully. The improvement from 72% to 92% is explicit, but the audit is based on a small manual sample. A plausible implication is that the dataset is materially cleaner than VoxBlink, but the paper does not present a large-scale label audit or a demographic error analysis (Lin et al., 2024).

3. Training regimes and speaker verification performance

VoxBlink2 is used primarily as a training corpus rather than as a standalone benchmark for conventional train/dev/test speaker verification splits. In the main speaker verification experiments, the training datasets are VoxBlink2 (VB2) and VoxCeleb2 (VC2), with evaluation on VoxCeleb1 test sets. Acoustic features are 80-dimensional log Mel-filterbank energies with 25 ms frame length, 10 ms hop size, and fixed input lengths of 200 or 500 frames. The speaker encoders include ResNet34, ResNet50, ResNet100, ResNet221, and ResNet293, with ASP, TSP, and SimAM variants (Lin et al., 2024).

The paper compares two training strategies. The first is a Mix-FT strategy: pre-train on VC2 and fine-tune on mixed data. The second is pre-train on VB2 and fine-tune on VC2. The latter is the stronger result. Using ResNet100-ASP + SimAM, the VC2 baseline gives 0.606 EER and 0.052 minDCF_{0.01} on VoxCeleb1-O, whereas VB2 pre-train + VC2 fine-tune gives 0.340 EER and 0.026 minDCF_{0.01}, which the paper describes as a 43.4% EER reduction relative to the VC2 baseline (Lin et al., 2024).

Scaling analysis is central to the paper’s interpretation of VoxBlink2. Random subsets with 5K, 10K, 30K, and the full 110K speakers show that more pretraining speakers consistently improve performance, and that larger models benefit more strongly from larger data. Model scaling from ResNet34 through ResNet293 also continues to improve performance, with no saturation reported in the tested range (Lin et al., 2024).

The final single-model result uses SimAM-ResNet293 with post-processing. The reported numbers on VoxCeleb1-O are 0.23 EER / 0.013 minDCF for the base model, 0.22 / 0.009 with AS-Norm, and 0.17 / 0.006 with QMF. The paper presents this as a new single-model state-of-the-art on VoxCeleb1-O (Lin et al., 2024).

These experiments also clarify what VoxBlink2 is not. It is not presented as a replacement for domain-aligned fine-tuning. Pretraining on VB2 alone yields 0.893 EER and 0.093 minDCF_{0.01} on VoxCeleb1-O, worse than VC2-only pretraining. The best results come from large-scale pretraining on VoxBlink2 followed by VC2 fine-tuning, not from naive dataset substitution (Lin et al., 2024).

4. Open-Set Speaker Identification benchmark

VoxBlink2 is also the basis for the paper’s Open-Set Speaker Identification agenda. OSSI is framed as a more realistic extension of verification: the system must match a probe utterance to a known gallery speaker or reject it as unknown. The benchmark is built from VoxBlink-clean, specifically 1,028,095 utterances from 18,381 speakers, and uses a gallery set $G$ together with a probe set $P = K \cup U$, where $K$ are known queries and $U$ are unknown queries (Lin et al., 2024).

Three protocol sizes are defined. VB-Eval-S uses 60 gallery speakers with 30 known and 30 unknown queries; VB-Eval-M uses 600 / 300 / 300; VB-Eval-L uses 6,000 / 3,000 / 3,000. Each protocol is further instantiated with 1, 3, or 5 enrollment utterances per speaker, yielding 9 evaluation protocols in total (Lin et al., 2024).

The paper formalizes the evaluation with DIR and FAR. For threshold $\theta$ and rank $n$,

$$\begin{aligned} DIR(\theta,n) = \frac{\lvert{ rank(p)\geq n \wedge sim(p,G_p)\geq \theta;p\in K}\rvert}{\lvert K \rvert} \end{aligned}$$

and

$$\begin{aligned} FAR(\theta) = \frac{\lvert{ sim(p,G_p)\geq \theta;p\in U}\rvert}{\lvert U \rvert}. \end{aligned}$$

The reported benchmark uses top-1 identification and reports DIR@FAR operating points (Lin et al., 2024).

The OSSI baseline uses a ResNet50 backbone pretrained on VB2; the paper notes that this model obtains 1.02% EER on VoxCeleb1-O. The open-set results show strong small-gallery performance but severe degradation at scale. For example, VB-Eval-S with 5 enrollments reaches 96.97 at DIR@FAR=0.001, whereas VB-Eval-L with 5 enrollments reaches only 24.94 at the same operating point. The general trends are explicit: more enrollment utterances improve DIR, but performance drops sharply as gallery size increases (Lin et al., 2024).

The OSSI benchmark is therefore one of VoxBlink2’s main conceptual contributions. It recasts large-scale speaker recognition as a 1:N plus rejection problem rather than only a 1:1 verification problem. The paper also compares speaker and face modalities and concludes that both perform strongly on small galleries but both degrade on VB-Eval-L, which it interprets as evidence that very-large-scale open-set identification remains unsolved (Lin et al., 2024).

5. Use in NIST SRE24 open-condition systems

Later work makes VoxBlink2 operationally important in NIST SRE24 open-condition systems. The ABC team describes VoxBlink2 as the central public corpus for the open condition, explicitly calling it the “largest publicly available speaker verification dataset” and using it to train ResNet-152-VB systems for conversational telephone speech after telephone-oriented adaptation. Their recipe includes downsampling the audio to 8kHz and applying GSM codec to 50% of the data via Sox to simulate the telephone channel, followed by fine-tuning on CTS Superset. With 40 s fine-tuning segments, the reported ResNet152-VB system reaches $C_{primary}=0.491$ and EER $=6.47\%$ on SRE24 eval (Barahona et al., 21 May 2025).

That paper is also explicit that VoxBlink2 is an open-condition-only resource. It is treated as a large-scale out-of-domain dataset for SRE24, and its value is tied to speaker diversity plus domain adaptation. The same paper notes that a hybrid recipe using original 16 kHz data plus a telephone-simulated copy improves broad-domain generalization relative to a purely telephone-adapted version (Barahona et al., 21 May 2025).

A second SRE24 example comes from the CL-UZH submission. There, VoxBlink2 appears only in the open-condition systems, not in the fixed condition. The open-condition description states that separate audio and visual networks were trained on VoxBlink2 and VoxCeleb/VoxCeleb2, with a ResNet293 audio model and a visual model described in prose as ResNet-based ArcFace, while the result table labels the visual system FaceNet. The multimodal system concatenates audio and visual embeddings and applies cosine distance (Farhadipour et al., 1 Oct 2025).

The CL-UZH numbers show the practical effect of VoxBlink2 as part of the open-condition training pool. In the open condition, ResNet293, Audio reports EER 10.2 / 8.47 on dev / eval, and ResNet+FN, AV reports EER 7.7 / 8.29 on dev / eval. The paper also states that the open-condition multimodal system “consistently outperform[ed] single-modality approaches,” especially in EER (Farhadipour et al., 1 Oct 2025).

These SRE24 papers are useful for clarifying a common misunderstanding. They do not describe VoxBlink2 as a special blink-dynamics resource, nor do they report blink-specific preprocessing. In both cases, VoxBlink2 functions as a training corpus for pretrained speaker or face encoders in open-condition systems (Barahona et al., 21 May 2025).

6. Limitations, interpretation, and data-governance issues

Despite its scale, VoxBlink2 has several explicit limitations. The paper does not provide a conventional train/dev/test split for VoxBlink2 itself; instead, VB2 is used for training, VoxCeleb1 is used for speaker verification evaluation, and VoxBlink-clean is used for OSSI. It also does not report demographic breakdowns such as gender balance, age, nationality, ethnicity, or detailed language proportions, even though the retrieval pipeline spans 18 languages and is described as globally diverse (Lin et al., 2024).

Quality estimation is also limited. The reported improvement from 72% to 92% label accuracy is based on manual inspection of only 50 randomly sampled speakers. This establishes a direction of improvement, but not a comprehensive error characterization (Lin et al., 2024).

The paper additionally implies several governance concerns. VoxBlink2 is mined from YouTube users, using both avatars and uploaded videos, which raises issues of consent, reuse of public personal data, biometric privacy, and misuse risk. These concerns are implicit in the data source and acknowledged in the paper’s caveat section, but there is no extended fairness audit, no detailed ethics section, and no exhaustive governance analysis (Lin et al., 2024).

From an evaluation standpoint, the largest OSSI protocol remains difficult. Even with 5 enrollment utterances per speaker, VB-Eval-L reaches only 24.94 at DIR@FAR=0.001. This suggests that VoxBlink2’s importance lies not only in improved verification pretraining but also in exposing an open-set identification regime that remains substantially unsolved (Lin et al., 2024).

Taken together, the available evidence supports a precise characterization. VoxBlink2 is a 100K+ speaker audio-visual corpus that substantially extends VoxBlink in scale, temporal coverage, and annotation purity; it supports strong speaker-verification pretraining when combined with domain-aligned fine-tuning; it introduces an open-set identification benchmark based on DIR@FAR; and it has already been adopted as a core open-condition training resource in SRE24 systems (Lin et al., 2024).