Exploring Speaker Diarization with Mixture of Experts

Gaobin Yang; Hang Chen; Jun Du; Maokui He; Ruoyu Wang; Shutong Niu

REVIEW 5 major objections 5 minor 37 references

Exploring Speaker Diarization with Mixture of Experts

T0 review · 5 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 A neural diarization system that adds a shared-and-soft mixture-of-experts layer to a memory-augmented sequence-to-sequence model reports the lowest diarization error rates on CHiME-6, DiPCo, and DIHARD-III, with a single model matching a…

desk verdict A real incremental gain from SS-MoE over the authors' own baseline, but the state-of-the-art claims are contradicted by the paper's own DIHARD-III table. read the letter →

arxiv 2506.14750 v1 pith:D6PE3LSE submitted 2025-06-17 cs.SD cs.AI

Gaobin Yang , Maokui He , Shutong Niu , Ruoyu Wang , Hang Chen , Jun Du This is my paper · ORCID

classification cs.SDcs.AI

keywords speakerdiarizationsequence-to-sequencemixtureofexpertsmemory-awaremulti-speakerembeddingCHiME-7DIHARD-IIIoverlappingspeechtarget-speakervoiceactivitydetection

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

Speaker diarization -- deciding who spoke when -- routinely fails when voices overlap. This paper argues that a neural system can get past that failure by retrieving one speaker embedding per known speaker from a memory module and feeding those embeddings, together with acoustic features, into a sequence-to-sequence decoder. The authors then add a Shared and Soft Mixture of Experts layer so the model can specialize to different acoustic conditions. Their resulting system, NSD-MS2S-SSMoE, reports the lowest diarization error rates on the CHiME-6 development and evaluation sets, the DiPCo development set, and most DIHARD-III domains, and the single model matches a six-checkpoint fusion of the no-MoE baseline.

What carries the argument

The carrying mechanism is the Shared and Soft Mixture of Experts (SS-MoE) module inserted into the decoder's feed-forward layers. Instead of hard routing, input tokens are softly dispatched to a fixed set of slots with a per-slot softmax, each collaborating expert processes a weighted average of all tokens, and the outputs are recombined with a per-token softmax; a shared expert processes every token in parallel and its output is added to the combined expert output. The module is initialized by copying the pretrained NSD-MS2S decoder parameters, with the feed-forward network replicated n+1 times to seed the experts.

What would settle it

Corrupt the speaker activity mask on the CHiME-6 development set -- delete one true speaker's mask or merge two speakers' masks -- and measure DER for NSD-MS2S-SSMoE; if the error rises sharply, the neural module inherits clustering mistakes rather than correcting them. A second decisive check replaces the SS-MoE experts with a single dense feed-forward layer of equal parameter count and asks whether the reported gains disappear.

Watch

Extended reading notes

Core claim

The central claim is that adding a Shared and Soft Mixture of Experts (SS-MoE) module inside the speaker-detection decoder of a memory-augmented sequence-to-sequence diarization system yields consistent accuracy gains on overlapping far-field speech. The system first uses clustering only to obtain a coarse speaker activity mask; the memory-aware multi-speaker embedding module then retrieves cleaner speaker embeddings, and the sequence-to-sequence decoder predicts frame-level speaker activity. On CHiME-6 the diarization error rate drops from 28.36% to 26.31% on the development set and from 29.45% to 28.51% on the evaluation set relative to the NSD-MS2S baseline. The single SS-MoE model matches a parameter-averaged fusion of six NSD-MS2S checkpoints, and on DIHARD-III it reports the best error rate in six of the eight evaluated domains.

Load-bearing premise

The reported gains assume that the initial clustering pipeline produces a good speaker activity mask and reliable i-vector seeds; if that clustering misses or splits a speaker, the memory retrieval and decoder are fed wrong speaker information, and the paper provides no ablation of how performance degrades with clustering quality.

Editorial extensions

If this is right

On CHiME-6, NSD-MS2S-SSMoE improves over NSD-MS2S by 7.23% relative DER on the development set and 3.19% on the evaluation set, and by 6.39% relative on the DiPCo development set.
A single NSD-MS2S-SSMoE model reaches the same CHiME-6 evaluation DER (28.51%) as the parameter-averaged fusion of six NSD-MS2S checkpoints.
On DIHARD-III, the system reports the best DER in broadcast news, courtroom, sociolinguistic lab, sociolinguistic field, conversational telephone speech, and meeting domains, but trails the QM-TS-VAD baseline in map task and clinical interviews.
Parameter transfer from a pretrained NSD-MS2S model lets SS-MoE converge more smoothly and reduces retraining cost by over 50%.
Increasing the number of experts from 2 to 6 improves results; beyond 6 the gains saturate, and the best decoder-layer placement differs by dataset (last three layers for CHiME-6, last two for DiPCo).

Reading between the lines

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

The absence of a matched-capacity dense control means the gains could come from added parameters rather than from the mixture-of-experts routing; a dense feed-forward network with the same parameter count would settle that.
Because the speaker activity mask and i-vector seeds come from a clustering pipeline, the method's advertised robustness to overlap is bounded by the clustering's recall; an ablation that perturbs the mask would show how much of the gain survives a miss.
The dataset-dependent optimal expert placement suggests that deploying this architecture to a new domain requires a per-dataset layer-selection search, which the paper does not provide a rule for.

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

Read the paper so you know what the actual contribution is before you see the abstract. The genuinely useful part is the within-family comparison: adding SS-MoE to their NSD-MS2S baseline gives consistent DER reductions on CHiME-6 dev (26.31 vs 28.36, 7.23% relative) and DiPCo dev (15.97 vs 17.06, 6.39%), and the single SS-MoE model roughly matches their six-checkpoint fusion on CHiME-6/DiPCo/Mixer 6 (Table V). That is a real, if incremental, engineering result, and the DIM ablation and convergence plots support it. The integration of MA-MSE with a Seq2Seq decoder plus a shared-expert variant of Soft MoE is a reasonable combination, and the parameter-transfer initialization trick is practical.

The problem is the headline. Contribution 4 and the abstract claim state-of-the-art on DIHARD-III and CHiME-7 first place. The paper's own Table IV contradicts the DIHARD-III claim: QM-TS-VAD has a better mean DER across the eight domains (roughly 7.8 vs 8.2) and better SOC.LAB (3.69 vs 3.85). The CHiME-7 first-place statement has no citation, and there's an unresolved '[?]' in the same intro paragraph. So the SOTA framing is not supported by the evidence presented.

Other soft spots are minor-to-moderate. Baselines are quoted from earlier papers without demonstrating matched VAD/scoring, so cross-system comparisons on CHiME-6/DiPCo are suggestive, not definitive. No error bars or significance tests; several 1-2% DER gaps could be noise. Expert count, slots, and layer placement are tuned per dataset on the dev set, which is a mild overfitting risk. The system depends on initial clustering for the mask and i-vectors, and there's no ablation of sensitivity to clustering quality. The manuscript is also unfinished: Fig.??, dangling sentences, and text-table mismatches (e.g., 17.51 vs 17.56; 1.09% vs 0.31%).

The narrow claim — SS-MoE over their own baseline — is plausible and internally consistent. The broad claim — SOTA — is not. If the authors correct the claims, add the CHiME-7 citation or drop the claim, and release code/data, the core result has lasting value. I'd send it to a serious referee with the expectation of major revision, but the referee should focus them on the claim-to-evidence gap rather than the architecture.

Referee Report

5 major / 5 minor

Summary. The paper proposes NSD-MS2S, a speaker diarization system that combines a memory-aware multi-speaker embedding (MA-MSE) module with a sequence-to-sequence decoder, and extends it with a Shared and Soft Mixture of Experts (SS-MoE) module. The experimental sections report DER/JER on CHiME-6, DiPCo, Mixer 6, and DIHARD-III, with internal comparisons against NSD-MS2S, NSD-MS2S-SoftMoE, and earlier TS-VAD baselines, plus ablations of the deep interactive module, expert count, expert placement, and parameter-transfer initialization. The abstract and Contribution 4 assert state-of-the-art results, including first place in the CHiME-7 main track and SOTA on the DIHARD-III evaluation set.

Significance. If the within-family comparisons are taken at face value, the paper gives a useful empirical demonstration that replacing feed-forward blocks with a shared-plus-soft MoE layer improves a memory-augmented sequence-to-sequence diarization system: NSD-MS2S-SSMoE consistently lowers DER/JER relative to NSD-MS2S on CHiME-6 dev/eval and DiPCo dev, and a single SS-MoE model roughly matches six-checkpoint parameter averaging on CHiME-6 eval. The DIM-vs-additive-attention ablation moves in the expected direction, and the parameter-transfer analysis quantifies a training-cost reduction. The study is empirical and does not ship code or leaderboard artifacts; the external SOTA claims are the weakest part of the evidence. The novelty over the authors' prior NSD-MA-MSE and NSD-MS2S work is incremental, but the measured within-family gains are a concrete contribution if the claims are appropriately scoped.

major comments (5)

[§I, Contribution 4; Table IV] The claim of state-of-the-art performance on the DIHARD-III evaluation set is not supported by the paper's own Table IV. Across the eight reported domains, QM-TS-VAD has a lower mean DER (roughly 7.8% versus 8.2% for NSD-MS2S-SSMoE) and wins on MAP TASK (1.59% vs. 5.14%), CLINICAL (9.81% vs. 11.48%), and SOC.LAB (3.69% vs. 3.85%). The text in §V.1 says the system achieves SOTA in SOC.LAB, which is contradicted by the table. Please either restrict the claim to the specific domains where the proposed system is actually best, or extend the comparison to the full official DIHARD-III leaderboard with matched scoring and matched VAD.
[§I, Contribution 4] The statement that NSD-MS2S achieved first place in the CHiME-7 main track is asserted without any citation, leaderboard reference, or comparison to official CHiME-7 results, and the introduction contains an unresolved '[?]' reference. Please provide verifiable challenge documentation, the official ranking, and the scoring conditions, or remove the first-place claim from the contributions and abstract.
[§IV.A.3] The DIHARD-III initialization removes silence segments 'based on official annotations.' If these are reference/ground-truth annotations, the proposed system enjoys an oracle VAD advantage that the published baselines in Table IV do not, making the DIHARD-III numbers not directly comparable. Please clarify what 'official annotations' means and, if oracle silence removal was used, re-run the experiments without it or report matched-VAD comparisons.
[§III.C and §IV.A.3] The MA-MSE module and the decoder initialization depend on a clustering-derived speaker activity mask S and seed i-vectors, yet the paper provides no experiment that varies clustering quality. Because the paper advertises robustness and generalization in challenging real-world scenarios, the sensitivity of the reported SS-MoE gains to initialization quality should be quantified, for example by corrupting or ablating the clustering input.
[Tables I–III] The abstract's broad 'state-of-the-art results' claim over CHiME-6, DiPCo, and Mixer 6 is not anchored to recent CHiME-7 submissions; the comparison tables include only older baselines such as x-vector clustering, DiaPer, EEND-VC, TS-VAD, and NSD-MA-MSE. Please either add current published systems from the CHiME-7 official benchmark or narrow the claim to 'improvement over the listed baselines'.

minor comments (5)

[Abstract and §I] The word 'spkeaker' appears in the abstract and again in the introduction; this typo should be corrected.
[§V.1] The text refers to 'incorporating sparse mixture-of-experts,' but SS-MoE is a shared and soft mixture of experts with dense, differentiable routing; the wording should be aligned with the method description.
[Figures 8–12] The figure axes and legends contain garbled '/uni000000...' placeholder strings; the figures need to be regenerated with readable labels.
[Figure 3] The figure caption contains untranslated Chinese characters ('分发模块结合模块'); the labels should be in English.
[§IV.A.4] The model is referred to as 'NSD-MS2S-MoE' at the start of the model configuration section, while the paper consistently uses 'NSD-MS2S-SSMoE' elsewhere; the naming should be unified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the SS-MoE gains are measured against external benchmarks, while the unsupported SOTA claims are an evidentiary issue, not a circularity issue.

full rationale

The paper's claimed derivation chain is not circular. The core new result, that NSD-MS2S-SSMoE improves over NSD-MS2S, is an empirical comparison on fixed test sets (Tables I-III): for example, CHiME-6 DER goes from 28.36%/29.45% for NSD-MS2S to 26.31%/28.51% for NSD-MS2S-SSMoE. These numbers are measured from trained models, not obtained by substituting the model's equations into its own outputs; no parameter is fitted to the evaluation metric and then renamed a prediction. The MA-MSE component is inherited from the authors' prior work [22], and the paper says so explicitly ('As outlined in [22], the core of the MA-MSE module is the memory component'), but this is ordinary incremental lineage rather than a self-referential proof: [22] is a separately published, externally reviewed system, and the paper also evaluates it as a baseline (NSD-MA-MSE) against independent clustering systems. The Soft MoE equations (1)-(2) are quoted from external work [27], and the SS-MoE extension (Eqs. 7-17) is a new architectural proposal tested on held-out data; there is no uniqueness theorem or ansatz imposed via self-citation. The headline SOTA and CHiME-7 first-place claims are under-supported: the first-place claim is uncited, and Table IV itself shows QM-TS-VAD achieving lower DER than NSD-MS2S-SSMoE on MAP TASK (1.59 vs 5.14) and SOC.LAB (3.69 vs 3.85). That is a correctness or evidence problem, not circularity. No equation in the paper reduces a predicted quantity to an input by construction, so the circularity score is 0.

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

The central claim rests on external inputs the paper does not verify: a clustering-based speaker mask and i-vectors, a memory of speaker basis vectors inherited from [22], simulated training data standing in for real conditions, and quoted baseline numbers from other papers. Architecture hyperparameters (expert count 6, slots 4, per-dataset layer placement, DIM depth 3) are set by dev-set sweeps; the fusion weights beta are trained. No new physical or conceptual entity is introduced; the expert modules are architectural components whose only evidence is the benchmark tables.

free parameters (6)

SS-MoE expert count n = 6
Hand-picked in Section IV.A.4; Fig. 11 shows gains saturate beyond 6 experts, so this is an operating point from a dev-set sweep, not a derived optimum.
SS-MoE slots per expert p = 4
Set in Section IV.A.4; no slot-count ablation is shown, so sensitivity is unknown.
SS-MoE layer placement = layers 4-6 for CHiME-6/DiPCo/Mixer 6; decoder layer 2 for DIHARD-III
Chosen per dataset via dev-set sweeps (Fig. 12); the dataset-specific choice indicates task-specific tuning that can overfit the dev set.
DIM depth = 3 blocks of 2 cross-attention layers
Fixed in Section III-C; no depth ablation is reported, so the choice is arbitrary within the paper.
Fusion weights beta_1, beta_2, beta_3 = learned scalars
Learnable interpolation weights in Eq. (3)-(4); they are trained from data rather than hand-set.
Memory basis-vector count K = not stated
The MA-MSE memory module and its basis-vector construction are inherited from [22]; K and the clustering recipe are not reported, leaving memory capacity uncontrolled.

assumptions (5)

domain assumption The clustering-based initialization (spectral clustering of x-vectors, or VBx for DIHARD-III) yields a speaker mask S and i-vectors accurate enough that errors do not cascade through MA-MSE frame selection and memory retrieval.
Section III-A and III-C: the mask S selects speaker-specific frames and the i-vector is concatenated into the aggregate embedding EA; the paper provides no sensitivity analysis to initialization errors.
domain assumption The memory basis vectors, cluster centers of speaker embeddings from additional datasets as in [22], generalize to the CHiME-6, DiPCo, Mixer 6, and DIHARD-III domains.
Section III-C: MA-MSE retrieves speaker embeddings from this memory; domain mismatch would degrade the retrieved embeddings. The construction is not re-validated here.
domain assumption Simulated multi-speaker conversations (jsalt2020 simulation plus pseudo-labeling) are a sufficient proxy for real far-field recordings.
Section IV.A.2: about 5,300 hours of CHiME-7 training data and 1,400 hours for DIHARD-III are mostly simulated; if the simulation gap is large, the reported gains may not transfer.
domain assumption Baseline numbers quoted from the cited papers are comparable to the authors' own runs, despite different VAD and initialization pipelines.
Section IV.A.5 and Tables I-IV: baselines are cited rather than re-run; the authors fine-tune the CHiME-7 VAD on CHiME-6+Mixer 6, which can shift all downstream numbers.
standard math The Soft MoE dispatch/combine equations (1)-(2) taken from [27] are correct and applicable inside the transformer decoder.
Section II.B: the routing mechanism is imported without re-derivation; the paper builds SS-MoE on top of it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring Speaker Diarization with Mixture of Experts." pith.science (2026). https://pith.science/paper/D6PE3LSE

@misc{pith2026250614750,
  author       = {Pith},
  title        = {Pith review of: Exploring Speaker Diarization with Mixture of Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D6PE3LSE}},
  note         = {Machine review of arXiv:2506.14750}
}

read the original abstract

In this paper, we propose a novel neural speaker diarization system using memory-aware multi-speaker embedding with sequence-to-sequence architecture (NSD-MS2S), which integrates a memory-aware multi-speaker embedding module with a sequence-to-sequence architecture. The system leverages a memory module to enhance speaker embeddings and employs a Seq2Seq framework to efficiently map acoustic features to speaker labels. Additionally, we explore the application of mixture of experts in speaker diarization, and introduce a Shared and Soft Mixture of Experts (SS-MoE) module to further mitigate model bias and enhance performance. Incorporating SS-MoE leads to the extended model NSD-MS2S-SSMoE. Experiments on multiple complex acoustic datasets, including CHiME-6, DiPCo, Mixer 6 and DIHARD-III evaluation sets, demonstrate meaningful improvements in robustness and generalization. The proposed methods achieve state-of-the-art results, showcasing their effectiveness in challenging real-world scenarios.

Figures

Figures reproduced from arXiv: 2506.14750 by the authors.

**Figure 2.** Soft MoE routing details. columns of the similarity matrix XΦ. Each slot representation is then obtained as a convex combination of all input tokens: Dij = exp((XΦ)ij ) Pm i ′=1 exp((XΦ)i ′j ) , X˜ = D⊤X (1) Each row of X˜ is routed to a designated expert based on its slot index. The expert function f⌊i/p⌋ then processes each slot independently to produce the intermediate output slots Y˜ i . A second softmax, applie… view at source ↗

**Figure 4.** Deep interactive module we identified that if the MA-MSE module structure is not optimized, it can severely affect performance in complex acoustic environments. Furthermore, overly simplistic mechanisms may limit the potential for performance improvement. To address this, we introduce the Deep Interactive Module (DIM), which replaces the additive attention mechanism with a dot-product attention mechanism and increa… view at source ↗

**Figure 5.** Illustration of the input dispatch module [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

Figures from the paper (6 more)

**Figure 7.** Figure 7: Illustration of the parameter transfer process from the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 9.** Figure 9: Detailed breakdown of DER components on the [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

**Figure 8.** Figure 8: Impact of the DIM module on the performance of the [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 12.** Figure 12: Impact of expert placement on system performance. [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]

**Figure 10.** Figure 10: Convergence comparison with different setups. [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]

**Figure 11.** Figure 11: Impact of the number of experts on system [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

[1]

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
[2]

Speaker diarization: A review of recent research,

X. Anguera, S. Bozonnet, N. Evans, C. Fredouille, G. Friedland, and O. Vinyals, “Speaker diarization: A review of recent research,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 20, no. 2, pp. 356–370, 2012

work page 2012
[3]

Unsupervised methods for speaker diarization: An integrated and iterative approach,

S. H. Shum, N. Dehak, R. Dehak, and J. R. Glass, “Unsupervised methods for speaker diarization: An integrated and iterative approach,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 21, no. 10, pp. 2015–2028, 2013

work page 2015
[4]

Front- end factor analysis for speaker verification,

N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front- end factor analysis for speaker verification,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 4, pp. 788–798, 2010

2010
[5]

Generalized end-to-end loss for speaker verification,

L. Wan, Q. Wang, A. Papir, and I. L. Moreno, “Generalized end-to-end loss for speaker verification,” inICASSP, 2018, pp. 4879–4883

work page 2018
[6]

X- vectors: Robust dnn embeddings for speaker recognition,

D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, “X- vectors: Robust dnn embeddings for speaker recognition,” inICASSP, 2018, pp. 5329–5333

work page 2018
[7]

A robust stopping criterion for agglomerative hierarchical clustering in a speaker diarization system,

K. J. Han and S. S. Narayanan, “A robust stopping criterion for agglomerative hierarchical clustering in a speaker diarization system,” in Eighth Annual Conference of the International Speech Communication Association, 2007

work page 2007
[8]

A study of the cosine distance-based mean shift for telephone speech diarization,

M. Senoussaoui, P. Kenny, T. Stafylakis, and P. Dumouchel, “A study of the cosine distance-based mean shift for telephone speech diarization,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 22, no. 1, pp. 217–227, 2013

work page 2013

Show all 37 references

[9]

Speaker diarization with LSTM,

Q. Wang, C. Downey, L. Wan, P. A. Mansfield, and I. L. Moreno, “Speaker diarization with LSTM,” inICASSP, 2018, pp. 5239–5243

2018
[10]

Speaker detection in the wild: Lessons learned from JSALT 2019,

P. Garc ´ıa, J. Villalba, H. Bredin, J. Du, D. Castan, A. Cristia, L. Bullock, L. Guo, K. Okabe, P. S. Nidadavoluet al., “Speaker detection in the wild: Lessons learned from JSALT 2019,” inOdyssey 2020 The Speaker and Language Recognition Workshop, 2020

2019
[11]

End-to-End Neural Speaker Diarization with Permutation-Free Objec- tives,

Y . Fujita, N. Kanda, S. Horiguchi, K. Nagamatsu, and S. Watanabe, “End-to-End Neural Speaker Diarization with Permutation-Free Objec- tives,”INTERSPEECH, pp. 4300–4304, 2019

2019
[12]

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

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

2019
[13]

End-to-End Speaker Diarization for an Unknown Number of Speakers with Encoder-Decoder Based Attractors,

S. Horiguchi, Y . Fujita, S. Watanabe, Y . Xue, and K. Nagamatsu, “End-to-End Speaker Diarization for an Unknown Number of Speakers with Encoder-Decoder Based Attractors,”INTERSPEECH, pp. 269–273, 2020

2020
[14]

Online end-to-end neural diarization with speaker-tracing buffer,

Y . Xue, S. Horiguchi, Y . Fujita, S. Watanabe, P. Garc ´ıa, and K. Naga- matsu, “Online end-to-end neural diarization with speaker-tracing buffer,” in2021 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2021, pp. 841–848

2021
[15]

Incorporating End-to-End Framework Into Target- Speaker V oice Activity Detection,

W. Wang and M. Li, “Incorporating End-to-End Framework Into Target- Speaker V oice Activity Detection,” inICASSP, 2022, pp. 8362–8366

2022
[16]

Target-Speaker V oice Activity Detection: A Novel Approach for Multi-Speaker Diarization in a Dinner Party Scenario,

I. Medennikov, M. Korenevsky, T. Prisyach, Y . Khokhlov, M. Ko- renevskaya, I. Sorokin, T. Timofeeva, A. Mitrofanov, A. Andrusenko, I. Podluzhnyet al., “Target-Speaker V oice Activity Detection: A Novel Approach for Multi-Speaker Diarization in a Dinner Party Scenario,” INTERS...

2020
[17]

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,”INTERSPEECH, pp. 3570–3574, 2021

2021
[18]

Chime-6 challenge: Tackling multispeaker speech recognition for unsegmented recordings,

S. Watanabe, M. Mandel, J. Barker, E. Vincent, A. Arora, X. Chang, S. Khudanpur, V . Manohar, D. Povey, D. Rajet al., “Chime-6 challenge: Tackling multispeaker speech recognition for unsegmented recordings,” inCHiME 2020-6th International Workshop on Speech Processing in Every...

2020
[19]

The chime- 7 dasr challenge: Distant meeting transcription with multiple devices in diverse scenarios,

S. Cornell, M. Wiesner, S. Watanabe, D. Raj, X. Chang, P. Garcia, Y . Masuyama, Z.-Q. Wang, S. Squartini, and S. Khudanpur, “The chime- 7 dasr challenge: Distant meeting transcription with multiple devices in diverse scenarios,”arXiv preprint arXiv:2306.13734, 2023

2023 arXiv
[20]

Target speaker voice activity detection with transformers and its integration with end- to-end neural diarization,

D. Wang, X. Xiao, N. Kanda, T. Yoshioka, and J. Wu, “Target speaker voice activity detection with transformers and its integration with end- to-end neural diarization,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, ...

2023
[21]

Target-Speaker V oice Activity Detection via Sequence-to-Sequence Prediction,

M. Cheng, W. Wang, Y . Zhang, X. Qin, and M. Li, “Target-Speaker V oice Activity Detection via Sequence-to-Sequence Prediction,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

2023
[22]

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
[23]

Dipco–dinner party corpus,

M. Van Segbroeck, A. Zaid, K. Kutsenko, C. Huerta, T. Nguyen, X. Luo, B. Hoffmeister, J. Trmal, M. Omologo, and R. Maas, “Dipco–dinner party corpus,” inInterSpeech, 2020, pp. 434–436

2020
[24]

Mixer 6,

L. Brandschain, D. Graff, C. Cieri, K. Walker, C. Caruso, and A. Neely, “Mixer 6,” inProceedings of the Seventh International Conference on Language Resources and Evaluation (LREC’10), 2010

2010
[25]

Bayesian HMM clustering of x-vector sequences (VBx) in speaker diarization: Theory, implementation and analysis on standard tasks,

F. Landini, J. Profant, M. Diez, and L. Burget, “Bayesian HMM clustering of x-vector sequences (VBx) in speaker diarization: Theory, implementation and analysis on standard tasks,”Computer Speech & Language, vol. 71, p. 101254, 2022

2022
[26]

Long short-term memory recurrent neural network architectures for large scale acoustic modeling,

H. Sak, A. W. Senior, and F. Beaufays, “Long short-term memory recurrent neural network architectures for large scale acoustic modeling,” inINTERSPEECH, 2014, pp. 338–342

2014
[27]

From sparse to soft mixtures of experts,

J. Puigcerver, C. Riquelme, B. Mustafa, and N. Houlsby, “From sparse to soft mixtures of experts,” pp. 1–33, 2024

2024
[28]

Dab-detr: Dynamic anchor boxes are better queries for detr,

L. Shilong, L. Feng, Z. Hao, Y . Xiao, Q. Xianbiao, S. Hang, Z. Jun, and L. Zhang, “Dab-detr: Dynamic anchor boxes are better queries for detr,” inICLR, 2022, pp. 10–20

2022
[29]

Semi-supervised multi-channel speaker diarization with cross- channel attention,

S. Wu, J. Du, M.-K. He, S. Niu, H. Chen, H. Tang, and C.-H. Lee, “Semi-supervised multi-channel speaker diarization with cross- channel attention,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2023, pp. 1–8

2023
[30]

Svitchboard 1: Small vocabulary tasks from switchboard 1,

S. King, C. Bartels, and J. Bilmes, “Svitchboard 1: Small vocabulary tasks from switchboard 1,” inInterspeech 2005-Eurospeech: 9th Euro- pean Conference on Speech Communication and Technology. Interna- tional Speech Communication Association, 2005, pp. 3385–3388

2005
[31]

The ami meeting corpus,

W. Kraaij, T. Hain, M. Lincoln, and W. Post, “The ami meeting corpus,” inProc. International Conference on Methods and Techniques in Behavioral Research, 2005, pp. 1–4

2005
[32]

Spot the conversation: speaker diarisation in the wild,

J. S. Chung, J. Huh, A. Nagrani, T. Afouras, and A. Zisserman, “Spot the conversation: speaker diarisation in the wild,” inInterSpeech, 2020, pp. 299–303

2020
[33]

Ecapa-tdnn: Em- phasized channel attention, propagation and aggregation in tdnn based speaker verification,

B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Em- phasized channel attention, propagation and aggregation in tdnn based speaker verification,” in21st Annual conference of the International Speech Communication Association (INTERSPEECH 2020). ISCA, 2020, pp. 3830–3834

2020
[34]

V oxceleb2: Deep speaker recognition,

J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep speaker recognition,” inINTERSPEECH. ISCA, 2018, pp. 1086–1090

2018
[35]

Bayesian hmm clustering of x-vector sequences (vbx) in speaker diarization: theory, implementa- tion and analysis on standard tasks,

F. Landini, J. Profant, M. Diez, and L. Burget, “Bayesian hmm clustering of x-vector sequences (vbx) in speaker diarization: theory, implementa- tion and analysis on standard tasks,”Computer Speech & Language, pp. 1–14, 2022

2022
[36]

Unsupervised adaptation with quality-aware masking to improve target-speaker voice activity detection for speaker diarization,

S. Niu, J. Du, M. He, C.-H. Lee, B. Li, and J. Li, “Unsupervised adaptation with quality-aware masking to improve target-speaker voice activity detection for speaker diarization,” inProc. INTERSPEECH, vol. 2023, 2023, pp. 3482–3486

2023
[37]

Ntt multi-speaker asr system for the dasr task of chime-8 challenge,

N. Kamo, N. Tawara, A. Ando, T. Kano, H. Sato, R. Ikeshita, T. Moriya, S. Horiguchi, K. Matsuura, A. Ogawaet al., “Ntt multi-speaker asr system for the dasr task of chime-8 challenge,”arXiv preprint arXiv:2409.05554, 2024. arXiv:2409.05554

2024 arXiv

Pith tools

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

[1] [1]

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

[2] [2]

Speaker diarization: A review of recent research,

X. Anguera, S. Bozonnet, N. Evans, C. Fredouille, G. Friedland, and O. Vinyals, “Speaker diarization: A review of recent research,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 20, no. 2, pp. 356–370, 2012

work page 2012

[3] [3]

Unsupervised methods for speaker diarization: An integrated and iterative approach,

S. H. Shum, N. Dehak, R. Dehak, and J. R. Glass, “Unsupervised methods for speaker diarization: An integrated and iterative approach,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 21, no. 10, pp. 2015–2028, 2013

work page 2015

[4] [4]

Front- end factor analysis for speaker verification,

N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front- end factor analysis for speaker verification,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 4, pp. 788–798, 2010

2010

[5] [5]

Generalized end-to-end loss for speaker verification,

L. Wan, Q. Wang, A. Papir, and I. L. Moreno, “Generalized end-to-end loss for speaker verification,” inICASSP, 2018, pp. 4879–4883

work page 2018

[6] [6]

X- vectors: Robust dnn embeddings for speaker recognition,

D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, “X- vectors: Robust dnn embeddings for speaker recognition,” inICASSP, 2018, pp. 5329–5333

work page 2018

[7] [7]

A robust stopping criterion for agglomerative hierarchical clustering in a speaker diarization system,

K. J. Han and S. S. Narayanan, “A robust stopping criterion for agglomerative hierarchical clustering in a speaker diarization system,” in Eighth Annual Conference of the International Speech Communication Association, 2007

work page 2007

[8] [8]

A study of the cosine distance-based mean shift for telephone speech diarization,

M. Senoussaoui, P. Kenny, T. Stafylakis, and P. Dumouchel, “A study of the cosine distance-based mean shift for telephone speech diarization,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 22, no. 1, pp. 217–227, 2013

work page 2013