Speaker Diarization with Overlapping Community Detection Using Graph Attention Networks and Label Propagation Algorithm

Jie Wang; Lin Li; Longjie Luo; Qingyang Hong; Wangjie Li; Xiaoxiao Li; Zhaoyang Li

REVIEW 3 major objections 5 minor 41 references

Speaker Diarization with Overlapping Community Detection Using Graph Attention Networks and Label Propagation Algorithm

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that replacing the clustering module of a speaker diarization system with a graph attention network followed by overlapping label propagation lowers diarization error to 15.94% on DIHARD-III without oracle voice activity…

desk verdict A plausible but incremental clustering pipeline whose headline SOTA number is currently unverifiable because the paper contradicts itself on a key hyperparameter and omits variance. read the letter →

arxiv 2506.02610 v1 pith:KSW3CG4O submitted 2025-06-03 cs.SD cs.AIeess.AS

Zhaoyang Li , Jie Wang , XiaoXiao Li , Wangjie Li , Longjie Luo , Lin Li , Qingyang Hong This is my paper

classification cs.SDcs.AIeess.AS

keywords speakerdiarizationgraphattentionnetworklabelpropagationalgorithmoverlappingcommunitydetectionDIHARDIIIerrorrateembeddingsPLDA

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a modular speaker diarization pipeline that treats the clustering step as overlapping community detection on a graph of speaker embeddings. A graph attention network first refines the embeddings and the pairwise similarity graph, then a label propagation algorithm assigns each node one or more speaker labels, so overlapping speech is handled in the clustering step itself rather than in post-processing. The authors report that the combined system achieves a 15.94% diarization error rate on the DIHARD-III evaluation set without oracle voice activity detection, which they identify as state of the art, and 11.07% with oracle VAD. The claim matters because overlapping speech and non-convex embedding distributions are the two main failure modes of standard clustering-based diarization, and the paper argues both can be addressed within a graph-based clustering module.

What carries the argument

The central machinery is a two-stage graph pipeline. First, a graph attention encoder (two GAT layers) maps 256-dimensional embeddings to a 64-dim space and a scoring module re-predicts edge weights, producing a refined affinity matrix that is fused with the original PLDA affinity matrix via a balance coefficient ε; the network is trained with binary cross-entropy against the ground-truth adjacency. Second, LPANNI label propagation iterates over this fused graph: nodes are sorted by node importance, labels are propagated with neighbor-node influence weights, and each node keeps all labels whose belonging coefficient exceeds a threshold, so a node at a community boundary receives multiple speaker labels. The path-length threshold β and iteration count τ control the propagation.

What would settle it

Rerun the OCDGALP pipeline on the DIHARD-III EVAL set with all hyperparameters chosen on the DEV set only, fixing β=2 as stated in Section 4.2 and then β=3 as suggested by Section 5.1, and compare the resulting DER values against the claimed 15.94% without oracle VAD and 11.07% with oracle VAD.

Watch

Extended reading notes

Core claim

The central claim is that replacing the clustering stage of a modular speaker diarization system with a graph-attention-refined overlapping label propagation algorithm (OCDGALP) reduces diarization error below previously published systems on DIHARD-III. On the evaluation split, the system reaches 15.94% DER without oracle VAD (best among the systems compared in Table 2) and 11.07% with oracle VAD. The authors attribute the gain to two mechanisms: the GAT's attention-weighted aggregation produces a cleaner affinity graph than the raw embeddings or a GCN, and the LPANNI label propagation assigns multiple labels to boundary nodes, which directly accounts for overlapping speakers rather than discarding those segments.

Load-bearing premise

The headline result holds only if the configuration that produced the reported numbers is the configuration the paper describes as optimal and if no evaluation-set labels were used to choose hyperparameters.

Editorial extensions

If this is right

If the result is correct, overlapping community detection can serve as a drop-in replacement for AHC or spectral clustering in modular diarization, moving overlap handling out of post-processing and into the clustering step.
The GAT's attention-weighted graph refinement is the source of improvement over raw PLDA and GCN-based refinement in the paper's own comparisons (S9 versus S6), so attention weighting appears to matter more than the graph architecture itself.
Without oracle VAD, the system achieves 15.94% DER, outperforming both clustering and end-to-end systems in Table 2, which suggests modular systems can still be state of the art when the clustering module is strong enough.
With oracle VAD, the clustering module alone reaches 11.07% DER, competitive with dedicated target-speaker VAD systems, indicating that explicit overlap modeling inside clustering can substitute for a separate overlap-handling front-end.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper does not state whether the hyperparameter sweep in Figure 3 was run on the DIHARD-III development or evaluation set; if evaluation labels were used for tuning, the reported 15.94% is not a clean out-of-sample measurement and should be verified on a held-out split.
The setup section fixes the propagation path length to β=2, while the hyperparameter study finds the lowest DER at β=3; this inconsistency means the published configuration may not match the configuration that produced the headline numbers, and a reproduction should test both values.
A natural extension is to evaluate the overlap-detection output directly, for example with segment-level overlap recall and precision, since an improved overall DER alone does not prove that the multiple labels assigned by LPANNI correspond to true overlapping speech regions.
The contribution of the GAT module likely depends on the 1,200-hour simulated training set; ablating the size or domain of that training data would isolate how much of the gain comes from the graph refinement versus the overlap-aware label propagation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The genuinely new thing here is an empirical result: 15.94% DER on DIHARD-III without oracle VAD, which edges out the listed systems. The method is a modular clustering pipeline — GAT to refine PLDA-based speaker embedding graphs, followed by LPANNI label propagation to detect overlapping communities. None of the components are new, so the novelty is limited to swapping GCN for GAT and using LPANNI in the overlap-aware community detection framework from Wang et al. [11]. That said, the benchmark outcome is new, and the paper is clearly written and easy to follow.

What the paper does well is the ablation structure. It includes the standard clustering baselines (AHC, SC, LPA), compares GCN and GAT across three clustering backends, and reports both oracle-VAD and no-oracle-VAD conditions on DEV and EVAL. The internal pattern is consistent: GAT beats GCN everywhere, and LPA-based overlap detection beats AHC/SC once the graph is refined. If those numbers hold, the contribution is meaningful for the modular diarization crowd.

The soft spot is load-bearing. Section 4.2 states the path length beta is set to 2; Section 5.1 says the lowest DER is obtained at beta=3 (Figure 3). Table 1 must have been generated with one of those values, and the paper never says which. If the final system used beta=3, Section 4.2 is wrong. If it used beta=2, the paper reports results from a configuration its own sweep shows is suboptimal. The text also never says whether the sweep in Figure 3 was run on DEV or EVAL. If EVAL was used for hyperparameter selection, the 15.94% figure is a post-selection statistic, not a clean out-of-sample measurement. The margin over EEND-M2F (16.07) is 0.13% DER — small enough that test-set tuning could plausibly produce it. The paper also reports 10-trial averages without standard deviations, so the reader cannot tell whether 0.13% exceeds run-to-run variance of the randomized label propagation. This is fixable: state the actual beta used, clarify the DEV/EVAL split for hyperparameter tuning, and report per-trial spread.

Who is this for: anyone tracking clustering-based speaker diarization or using DIHARD-III as a benchmark. The GCN-to-GAT comparison is a fair data point, and the LPANNI adaptation is worth knowing about. But the paper should not be accepted as-is with the internal contradiction unresolved.

Recommendation: send it to peer review, but expect a revision. The referee should ask for the missing configuration details and variance before the SOTA claim can be trusted. If the numbers survive that scrutiny, it is a solid incremental contribution.

Referee Report

3 major / 5 minor

Summary. The paper proposes OCDGALP, a modular speaker diarization system that builds a PLDA-based affinity graph from segment-level speaker embeddings, refines the graph with a two-layer graph attention network, fuses the reconstructed affinity matrix with the original one, and then applies an LPANNI-style label propagation algorithm to perform clustering and overlapping-speech detection simultaneously. The method is evaluated on DIHARD-III under a 0 ms collar with and without oracle VAD. The headline results are a 15.94% DER without oracle VAD and 11.07% with oracle VAD on the evaluation set, which the paper claims is state-of-the-art in the no-oracle-VAD condition.

Significance. If the reported numbers survive scrutiny, the paper would show that a modular clustering-based system can outperform recent end-to-end and fusion systems on DIHARD-III in the no-oracle-VAD condition, which is a useful and nontrivial result. The pipeline is described clearly at the conceptual level, the evaluation is against an external public benchmark rather than a fitted objective, and the comparison table is informative. However, the central claim is currently not verifiable from the manuscript because the experimental setup text and the hyperparameter sweep disagree on the value of beta, and because the paper does not disclose whether the sweep was performed on the evaluation set. The missing variance statistics are also material, since the claimed advantage over the next-best system is only 0.13% absolute DER.

major comments (3)

[Section 4.2 vs Section 5.1] Section 4.2 states "we set the maximum number of iterations τ to 80 and the path length β to 2," but Section 5.1 reports that the lowest DER is obtained when ε = 0.5 and β = 3. The manuscript never states which configuration produced the S9 row of Table 1. If Table 1 used β = 3, the setup description is wrong; if it used β = 2, the headline 15.94% DER was produced by a configuration that the paper's own Figure 3 identifies as suboptimal. Please state the exact hyperparameters used for Table 1 and reconcile the text with the reported experiments.
[Section 5.1] Figure 3 is described only as showing results "on the DIHARD III dataset," without specifying whether the sweep over ε and β was run on the DEV set, the EVAL set, or both. The central claim is an EVAL number, and if the hyperparameters were selected by evaluating EVAL, the 15.94% DER is a post-selection statistic rather than a clean out-of-sample measurement. The manuscript must state the split used for hyperparameter selection and, if EVAL was used, either re-run the selection on DEV only or re-evaluate with a properly held-out protocol.
[Section 4.2 and Table 1] The paper states that each experimental outcome is averaged over 10 repeated trials, but no standard deviations, confidence intervals, or per-trial results are reported. This is important because label propagation is randomized and the S9 no-oracle Eval DER of 15.94% is only 0.13% absolute below the EEND-M2F figure of 16.07% in Table 2. Without variance information, the reader cannot judge whether the reported advantage is larger than run-to-run noise. Please report the spread of the 10 trials and, preferably, a paired significance test against the relevant baselines.

minor comments (5)

[Section 5.3] The sentence claiming that with oracle VAD the system "matched the performance of the best system reported to date" is contradicted by Table 2, which lists Seq2Seq-TSVAD at 10.77% DER, lower than the reported 11.07% for OCDGALP. This wording should be corrected, even though it does not affect the no-oracle-VAD SOTA claim.
[Figure 3] The legend in Figure 3 appears to use α for the path-length parameter, while the text and caption use β; please unify the notation.
[Section 4.2] The GAT training hyperparameters (learning rate, number of epochs, batch size, early stopping, and the number of simulated graphs used for training) are not reported, which limits reproducibility of the graph refinement module.
[Table 1] The acronym LPA-OCD is used in Table 1 but not defined in the text; please define it at first use, for example when introducing S6.
[Abstract and Introduction] The phrase "state-of-the-art" should be qualified as "state-of-the-art among the systems listed in Table 2" or "to the best of our knowledge," since the claim depends on the comparison set and on the unresolved hyperparameter-selection issue.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the DER benchmark is an external empirical measurement; the main caveat is an internal beta inconsistency, which is a soundness/reproducibility issue rather than a circular derivation.

full rationale

The paper's central claim (15.94% DER on DIHARD-III EVAL without oracle VAD) is an empirical result against an external benchmark. It is not defined in terms of the method's own fitted quantities: the GAT is trained on simulated mixtures with ground-truth adjacency, and the LPANNI clustering is cited from external work [19]. The graph construction (PLDA + threshold) and fusion rule (Eq. 6) are described explicitly and do not encode the target DER. The method builds on the authors' prior GCN community-detection work [11], but [11] is used as a baseline (S4-S6), not as a proof or uniqueness argument, so the main claim has independent evidential content. The only load-bearing concern is methodological, not circular: Section 4.2 fixes beta = 2 while Section 5.1 reports the lowest DER at beta = 3, and Section 5.1 does not specify whether the sweep was run on DEV or EVAL. If the final EVAL number came from a beta/epsilon selection on EVAL, the SOTA claim would be a post-selection statistic rather than an out-of-sample prediction; if it used beta = 2, the described setup is suboptimal per the paper's own Figure 3. This is a missing-support/reproducibility flaw, not a reduction of the result to its inputs by construction. Accordingly, no circular step is identified and the circularity score is low.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities, particles, forces, or conserved quantities. It is an algorithmic combination of existing graph neural network and label propagation techniques.

free parameters (6)

mu (edge retention threshold) = 0.3
Edges with PLDA similarity below this threshold are removed; the value is set by hand with no ablation shown.
epsilon (affinity fusion coefficient) = 0.5
Chosen from hyperparameter analysis in Section 5.1; balances reconstructed and original affinity matrices.
beta (path length threshold) = 2 in Section 4.2, but 3 reported optimal in Section 5.1
Controls the path length used in computing node similarity for label propagation; the paper is internally inconsistent about the value actually used.
tau (max label propagation iterations) = 80
Maximum label propagation iterations; set by hand.
GAT hidden layer widths = 256-128-64-64-1
Architecture dimensions chosen by hand; no ablation is shown for these widths.
segment length and shift = 1.5s / 0.75s
Segmentation choices fixed by hand and not ablated.

assumptions (5)

standard math Graph attention equations and attention coefficient computation from Velickovic et al. are correct and applicable to speaker embedding graphs.
Used in Sections 2.1 and 3.2; accepted prior work.
domain assumption PLDA similarity scores provide a valid affinity measure between segment-level speaker embeddings.
Used in graph construction in Section 3.1; no validation on DIHARD-III is provided.
domain assumption A node (segment embedding) may belong to multiple speaker communities, and these community memberships correspond to actual overlapping speakers.
Central to Section 3.3: nodes with multiple speaker labels are considered overlapping nodes. This assumes a single 1.5s segment embedding can represent multiple simultaneous speakers.
domain assumption The simulated 2 to 9 speaker mixtures from VoxCeleb1, VoxCeleb2, and LibriSpeech are representative of DIHARD-III acoustic conditions for GAT training.
Used in Section 4.1; no domain adaptation is described.
ad hoc to paper Stability of the label set and dominant labels is a sufficient convergence criterion for the overlapping community detection.
Defined in Section 3.3; no formal guarantee of convergence to a meaningful partition is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Speaker Diarization with Overlapping Community Detection Using Graph Attention Networks and Label Propagation Algorithm." pith.science (2026). https://pith.science/paper/KSW3CG4O

@misc{pith2026250602610,
  author       = {Pith},
  title        = {Pith review of: Speaker Diarization with Overlapping Community Detection Using Graph Attention Networks and Label Propagation Algorithm},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KSW3CG4O}},
  note         = {Machine review of arXiv:2506.02610}
}

read the original abstract

In speaker diarization, traditional clustering-based methods remain widely used in real-world applications. However, these methods struggle with the complex distribution of speaker embeddings and overlapping speech segments. To address these limitations, we propose an Overlapping Community Detection method based on Graph Attention networks and the Label Propagation Algorithm (OCDGALP). The proposed framework comprises two key components: (1) a graph attention network that refines speaker embeddings and node connections by aggregating information from neighboring nodes, and (2) a label propagation algorithm that assigns multiple community labels to each node, enabling simultaneous clustering and overlapping community detection. Experimental results show that the proposed method significantly reduces the Diarization Error Rate (DER), achieving a state-of-the-art 15.94% DER on the DIHARD-III dataset without oracle Voice Activity Detection (VAD), and an impressive 11.07% with oracle VAD.

Figures

Figures reproduced from arXiv: 2506.02610 by the authors.

**Figure 1.** An illustration of the speaker diarization system pipeline with the OCDGALP clustering method. can impact the classification performance of these systems. For high-dimensional speaker embeddings such as x-vector [9], Graph Neural Network (GNN) can effectively leverage relational information between data points to further refine the speaker embeddings, thereby enhancing their representativeness. Wang et al. [10] in… view at source ↗

**Figure 2.** The architecture of the Overlapping Community Detection based on Graph Attention and Label Propagation Algorithm. where ε ∈ [0, 1] is a balance coefficient that controls the influence of different affinity matrices. The loss function used for model training is defined as the difference between the fused affinity matrix and the ground truth adjacency matrix, calculated using Binary Cross-Entropy (BCE) loss. 3.3. Ov… view at source ↗

**Figure 3.** The comparison of DER(%) with different ε and β on the DIHARD III dataset. S6 integrate GCN modules, while S7 to S9 incorporate GAT modules into the baseline systems. Notably, S9 represents our proposed OCDGALP system. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 35 canonical work pages

[11]

Acknowledgement This work was supported in part by the National Natural Science Foundation of China under Grants 62371407 and 62276220, and the Innovation of Policing Science and Tech- nology, Fujian province (Grant number: 2024Y0068)

work page
[1]

Speaker Diarization with Overlapping Community Detection Using Graph Attention Networks and Label Propagation Algorithm

Introduction Speaker diarization is designed to segment the conversational audio signals into distinct segments with labeled speakers’ iden- tities, effectively addressing the problem of “who spoke when” [1]. In recent years, speaker diarization technology has gained widespread application across diverse fields, such as meeting transcription, conversation...

work page Pith review arXiv 2025
[2]

Graph attention network In our work, a modified GAT model is used to update the speaker embedding graph

Related works 2.1. Graph attention network In our work, a modified GAT model is used to update the speaker embedding graph. Assuming L-layer graph attention network, where l ∈ {1, 2, ..., L} represents the layer index, we take the embedding matrix H ∈ RK×D, which consists of pre- extracted speaker embeddings, along with the adjacency matrix A ∈ RK×K as th...

work page
[3]

For a given network G = ( V, E), where V represents the set of nodes and E represents the set of edges, and u ∈ V , v ∈ N g(u), and N g(u) represents the set of neighbors of u, we introduce additional definitions for nodes in the graph: Node Importance (NI (u)), Node Similarity (Sim(u, v)), and Neighbor Node Influence ( NNI v(u)), which are defined in the...

work page
[4]

Nodes are sorted according to NI (u) for label updating. At this stage, nodes will receive multiple dominant labels from neighboring nodes, forming a label set: LN g= {l(c1, b1), l(c2, b2), ...l(cv, bv)}, (3) where l(cv, bv) denotes the dominant label of neighbor node v, and bv denotes the belonging coefficient of nodev to com- munity cv. The dominant lab...

work page
[5]

Next, labels satisfying b′(c, u) < 1/|L′| are adaptively removed, forming the refined label set L′′

Recalculate the new belonging coefficient b′(c, u) of node u to community c using NNI and LN g: b′(c, u) = P l(cv ,bv )∈LN g,v∈N g(u),cv =c b(cv, v)N N Iv(u) P l(cv ,bv )∈LN g,v∈N g(u) b(cv, v)N N Iv(u) (4) Then, the updated label set L′ of node u is generated: L′ = {l(c1, b′ 1), l(c2, b′ 2), ...l(c|L′|, b′ |L′|)}, (5) where |L′| denotes the number of the...

work page
[6]

The update process follows a historical label priority strat- egy to reduce propagation randomness

Normalize the belonging coefficients of the labels in L′′ to obtain the final label set Lu. The update process follows a historical label priority strat- egy to reduce propagation randomness. If multiple labels share the highest belonging coefficients, the dominant label is cho- sen from those present in the previous iteration; otherwise, it is selected r...

work page
[7]

The following sections provide a detailed introduction to each module of OCDGALP

Proposed Method The overall framework of OCDGALP is illustrated in Figure 2. The following sections provide a detailed introduction to each module of OCDGALP. 3.1. Graph Generation In a graph G = (V, E, A) composed of conversations, as illus- trated in Figure 2(a), the nodes V = {v1, v2, . . . , vN } ∈RN are derived from speaker embeddings extracted from ...

work page

Show all 41 references

[8]

Datasets and experimental setup 4.1. Evaluation and training data preparation We evaluated our speaker diarization system on the DIHARD- III corpus, which includes a 34.15-hour development (DEV) set and a 33.01-hour evaluation (EV AL) set. The training sets for different modul...
[9]

Choice of hyperparameters We investigate the effect of the balance coefficient ε and the path length threshold β on the OCDGALP system

Experimental results 5.1. Choice of hyperparameters We investigate the effect of the balance coefficient ε and the path length threshold β on the OCDGALP system. As shown in Figure 3, the lowest DER is obtained when ε = 0 .5 and β = 3 . This result demonstrates the effectivene...
[10]

Our primary goal is to leverage the topological struc- ture of speaker embeddings and fully exploit graph-based infor- mation to achieve both clustering and overlap detection

Conclusion In this paper, we propose the OCDGALP model for speaker di- arization. Our primary goal is to leverage the topological struc- ture of speaker embeddings and fully exploit graph-based infor- mation to achieve both clustering and overlap detection. To this end, we inc...
[12]

A review of speaker diarization: Recent advances with deep learning,

T. J. Park, N. Kanda, D. Dimitriadis, K. J. Han, S. Watanabe, and S. Narayanan, “A review of speaker diarization: Recent advances with deep learning,” Computer Speech & Language , vol. 72, p. 101317, 2022

2022
[13]

Artificial neural network features for speaker diarization,

S. H. Yella, A. Stolcke, and M. Slaney, “Artificial neural network features for speaker diarization,” in 2014 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2014, pp. 402–406

2014
[14]

End-to-end neural speaker diarization with self- attention,

Y . Fujita, N. Kanda, S. Horiguchi, Y . Xue, K. Nagamatsu, and S. Watanabe, “End-to-end neural speaker diarization with self- attention,” in 2019 IEEE Automatic Speech Recognition and Un- derstanding Workshop (ASRU). IEEE, 2019, pp. 296–303

2019
[15]

Neural speaker diarization with speaker-wise chain rule,

Y . Fujita, S. Watanabe, S. Horiguchi, Y . Xue, J. Shi, and K. Naga- matsu, “Neural speaker diarization with speaker-wise chain rule,” arXiv preprint arXiv:2006.01796, 2020

2006 arXiv
[16]

Ansd-ma-mse: Adap- tive neural speaker diarization using memory-aware multi-speaker embedding,

M.-K. He, J. Du, Q.-F. Liu, and C.-H. Lee, “Ansd-ma-mse: Adap- tive neural speaker diarization using memory-aware multi-speaker embedding,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 1561–1573, 2023

2023
[17]

Target- speaker voice activity detection via sequence-to-sequence predic- tion,

M. Cheng, W. Wang, Y . Zhang, X. Qin, and M. Li, “Target- speaker voice activity detection via sequence-to-sequence predic- tion,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

2023
[18]

Priors for speaker counting and diarization with ahc

G. Sell, A. McCree, and D. Garcia-Romero, “Priors for speaker counting and diarization with ahc.” in InterSpeech, 2016, pp. 2194–2198

2016
[19]

Similarity measure- ment of segment-level speaker embeddings in speaker diariza- tion,

W. Wang, Q. Lin, D. Cai, and M. Li, “Similarity measure- ment of segment-level speaker embeddings in speaker diariza- tion,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 2645–2658, 2022

2022
[20]

Speaker diarization with lstm,

Q. Wang, C. Downey, L. Wan, P. A. Mansfield, and I. L. Moreno, “Speaker diarization with lstm,” in 2018 IEEE International con- ference on acoustics, speech and signal processing (ICASSP) . IEEE, 2018, pp. 5239–5243

2018
[21]

Speaker diarization with session-level speaker em- bedding refinement using graph neural networks,

J. Wang, X. Xiao, J. Wu, R. Ramamurthy, F. Rudzicz, and M. Brudno, “Speaker diarization with session-level speaker em- bedding refinement using graph neural networks,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, ...

2020
[22]

Community de- tection graph convolutional network for overlap-aware speaker di- arization,

J. Wang, Z. Chen, H. Zhou, L. Li, and Q. Hong, “Community de- tection graph convolutional network for overlap-aware speaker di- arization,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

2023
[23]

Reformulating speaker diarization as community detection with emphasis on topological structure,

S. Zheng and H. Suo, “Reformulating speaker diarization as community detection with emphasis on topological structure,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 8097– 8101

2022
[24]

The third dihard diarization challenge,

N. Ryant, P. Singh, V . Krishnamohan, R. Varma, K. Church, C. Cieri, J. Du, S. Ganapathy, and M. Liberman, “The third dihard diarization challenge,” arXiv preprint arXiv:2012.01477, 2020

2012 arXiv
[25]

Graph attention networks,

P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y . Bengio et al. , “Graph attention networks,” stat, vol. 1050, no. 20, pp. 10–48 550, 2017

2017
[26]

Near linear time algorithm to detect community structures in large-scale net- works,

U. N. Raghavan, R. Albert, and S. Kumara, “Near linear time algorithm to detect community structures in large-scale net- works,” Physical Review E—Statistical, Nonlinear , and Soft Mat- ter Physics, vol. 76, no. 3, p. 036106, 2007

2007
[27]

Finding overlapping communities in networks by label propagation,

S. Gregory, “Finding overlapping communities in networks by label propagation,” New journal of Physics , vol. 12, no. 10, p. 103018, 2010

2010
[28]

Slpa: Uncovering overlap- ping communities in social networks via a speaker-listener inter- action dynamic process,

J. Xie, B. K. Szymanski, and X. Liu, “Slpa: Uncovering overlap- ping communities in social networks via a speaker-listener inter- action dynamic process,” in 2011 ieee 11th international confer- ence on data mining workshops . IEEE, 2011, pp. 344–349

2011
[29]

Detecting overlapping communities in networks via dominant la- bel propagation,

H.-L. Sun, J.-B. Huang, Y .-Q. Tian, Q.-B. Song, and H.-L. Liu, “Detecting overlapping communities in networks via dominant la- bel propagation,” Chinese Physics B , vol. 24, no. 1, p. 018703, 2015

2015
[30]

Lpanni: Overlapping com- munity detection using label propagation in large-scale complex networks,

M. Lu, Z. Zhang, Z. Qu, and Y . Kang, “Lpanni: Overlapping com- munity detection using label propagation in large-scale complex networks,” IEEE Transactions on Knowledge and Data Engineer- ing, vol. 31, no. 9, pp. 1736–1749, 2018

2018
[31]

Probabilistic linear discriminant analysis,

S. Ioffe, “Probabilistic linear discriminant analysis,” in Computer Vision–ECCV 2006: 9th European Conference on Computer Vi- sion, Graz, Austria, May 7-13, 2006, Proceedings, Part IV 9 . Springer, 2006, pp. 531–542

2006
[32]

Powerset multi-class cross en- tropy loss for neural speaker diarization,

A. Plaquet and H. Bredin, “Powerset multi-class cross en- tropy loss for neural speaker diarization,” arXiv preprint arXiv:2310.13025, 2023

2023 arXiv
[33]

Asv- subtools: Open source toolkit for automatic speaker verification,

F. Tong, M. Zhao, J. Zhou, H. Lu, Z. Li, L. Li, and Q. Hong, “Asv- subtools: Open source toolkit for automatic speaker verification,” in ICASSP 2021-2021 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) . IEEE, 2021, pp. 6184–6188

2021
[34]

Encoder-decoder based attractors for end-to-end neural diariza- tion,

S. Horiguchi, Y . Fujita, S. Watanabe, Y . Xue, and P. Garcia, “Encoder-decoder based attractors for end-to-end neural diariza- tion,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 1493–1507, 2022

2022
[35]

The hitachi- jhu dihard iii system: Competitive end-to-end neural diarization and x-vector clustering systems combined by dover-lap,

S. Horiguchi, N. Yalta, P. Garcia, Y . Takashima, Y . Xue, D. Raj, Z. Huang, Y . Fujita, S. Watanabe, and S. Khudanpur, “The hitachi- jhu dihard iii system: Competitive end-to-end neural diarization and x-vector clustering systems combined by dover-lap,” arXiv preprint arXiv:2...

2021 arXiv
[36]

Ustc-nelslip system description for dihard-iii chal- lenge,

Y . Wang, M. He, S. Niu, L. Sun, T. Gao, X. Fang, J. Pan, J. Du, and C.-H. Lee, “Ustc-nelslip system description for dihard-iii chal- lenge,” arXiv preprint arXiv:2103.10661, 2021

2021 arXiv
[37]

But system description for the third dihard speech diarization challenge,

F. Landini, A. Lozano-Diez, L. Burget, M. Diez, A. Silnova, K. Zmolıkov ´a, O. Glembek, P. Matejka, T. Stafylakis, and N. Br¨ummer, “But system description for the third dihard speech diarization challenge,” in Proc. 3rd DIHARD Speech Diarization Challenge Workshop, 2021

2021
[38]

Diaper: End-to-end neural diarization with perceiver-based attractors,

F. Landini, T. Stafylakis, L. Burget et al. , “Diaper: End-to-end neural diarization with perceiver-based attractors,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2024

2024
[39]

Online neural diarization of unlimited numbers of speakers using global and local attractors,

S. Horiguchi, S. Watanabe, P. Garc ´ıa, Y . Takashima, and Y . Kawaguchi, “Online neural diarization of unlimited numbers of speakers using global and local attractors,” IEEE/ACM Trans- actions on Audio, Speech, and Language Processing , vol. 31, pp. 706–720, 2022

2022
[40]

End-to-end speaker segmen- tation for overlap-aware resegmentation,

H. Bredin and A. Laurent, “End-to-end speaker segmen- tation for overlap-aware resegmentation,” arXiv preprint arXiv:2104.04045, 2021

2021 arXiv
[41]

Eend- m2f: Masked-attention mask transformers for speaker diariza- tion,

M. H ¨ark¨onen, S. J. Broughton, and L. Samarakoon, “Eend- m2f: Masked-attention mask transformers for speaker diariza- tion,” arXiv preprint arXiv:2401.12600, 2024

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [11]

Acknowledgement This work was supported in part by the National Natural Science Foundation of China under Grants 62371407 and 62276220, and the Innovation of Policing Science and Tech- nology, Fujian province (Grant number: 2024Y0068)

work page

[2] [1]

Speaker Diarization with Overlapping Community Detection Using Graph Attention Networks and Label Propagation Algorithm

Introduction Speaker diarization is designed to segment the conversational audio signals into distinct segments with labeled speakers’ iden- tities, effectively addressing the problem of “who spoke when” [1]. In recent years, speaker diarization technology has gained widespread application across diverse fields, such as meeting transcription, conversation...

work page Pith review arXiv 2025

[3] [2]

Graph attention network In our work, a modified GAT model is used to update the speaker embedding graph

Related works 2.1. Graph attention network In our work, a modified GAT model is used to update the speaker embedding graph. Assuming L-layer graph attention network, where l ∈ {1, 2, ..., L} represents the layer index, we take the embedding matrix H ∈ RK×D, which consists of pre- extracted speaker embeddings, along with the adjacency matrix A ∈ RK×K as th...

work page

[4] [3]

For a given network G = ( V, E), where V represents the set of nodes and E represents the set of edges, and u ∈ V , v ∈ N g(u), and N g(u) represents the set of neighbors of u, we introduce additional definitions for nodes in the graph: Node Importance (NI (u)), Node Similarity (Sim(u, v)), and Neighbor Node Influence ( NNI v(u)), which are defined in the...

work page

[5] [4]

Nodes are sorted according to NI (u) for label updating. At this stage, nodes will receive multiple dominant labels from neighboring nodes, forming a label set: LN g= {l(c1, b1), l(c2, b2), ...l(cv, bv)}, (3) where l(cv, bv) denotes the dominant label of neighbor node v, and bv denotes the belonging coefficient of nodev to com- munity cv. The dominant lab...

work page

[6] [5]

Next, labels satisfying b′(c, u) < 1/|L′| are adaptively removed, forming the refined label set L′′

Recalculate the new belonging coefficient b′(c, u) of node u to community c using NNI and LN g: b′(c, u) = P l(cv ,bv )∈LN g,v∈N g(u),cv =c b(cv, v)N N Iv(u) P l(cv ,bv )∈LN g,v∈N g(u) b(cv, v)N N Iv(u) (4) Then, the updated label set L′ of node u is generated: L′ = {l(c1, b′ 1), l(c2, b′ 2), ...l(c|L′|, b′ |L′|)}, (5) where |L′| denotes the number of the...

work page

[7] [6]

The update process follows a historical label priority strat- egy to reduce propagation randomness

Normalize the belonging coefficients of the labels in L′′ to obtain the final label set Lu. The update process follows a historical label priority strat- egy to reduce propagation randomness. If multiple labels share the highest belonging coefficients, the dominant label is cho- sen from those present in the previous iteration; otherwise, it is selected r...

work page

[8] [7]

The following sections provide a detailed introduction to each module of OCDGALP

Proposed Method The overall framework of OCDGALP is illustrated in Figure 2. The following sections provide a detailed introduction to each module of OCDGALP. 3.1. Graph Generation In a graph G = (V, E, A) composed of conversations, as illus- trated in Figure 2(a), the nodes V = {v1, v2, . . . , vN } ∈RN are derived from speaker embeddings extracted from ...

work page