Pith. sign in

REVIEW 5 major objections 5 minor 46 references

End-to-End Diarization utilizing Attractor Deep Clustering

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 Diarization error falls to 4.99% through attractor deep clustering.

desk verdict A compact EEND variant with a genuinely new DPCL-style loss that reports 4.99% DER on CALLHOME, but the gain over published baselines isn't attribution-safe until they're retrained under matched data. read the letter →

arxiv 2506.11090 v1 pith:SVLDCLJV submitted 2025-06-05 cs.SD eess.AS

David Palzer , Matthew Maciejewski , Eric Fosler-Lussier This is my paper
classification cs.SDeess.AS
keywords speakerdiarizationend-to-endneuralattractorsdeepclusteringconformeranglelossorthogonalityconstraintspermutationinvarianttraining
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 tries to establish that an end-to-end neural diarization system can reach a very low diarization error rate while staying compact, by combining conformer decoders, transformer-updated attractors, and a deep clustering angle loss. On the CALLHOME test set, the proposed system reports a diarization error rate of 4.99% with only 15.3 million parameters, lower than the published rates of several larger baseline systems. The central claim is that extending deep clustering so that label vectors are built from speaker attractors, and adding orthogonality constraints, produces a structured embedding geometry that reduces false alarms and speaker confusion. This matters because it suggests that structured representation losses can make end-to-end diarization both more accurate and more practical for resource-constrained use.

What carries the argument

The central object is the speaker attractor, a learned 256-dimensional prototype vector initialized as an 8-by-256 matrix and refined layer by layer by a transformer decoder that first applies self-attention among attractors and then cross-attention to the conformer output. The new auxiliary loss, A-DPCL, forms a label-attractor vector for each time frame by multiplying the frame's speaker-activity labels by the attractor matrix and L2-normalizing, then compares pairwise inner products of these label-attractor vectors with pairwise inner products of the audio embeddings. This mechanism directly couples the geometric directions of attractors and audio embeddings, while orthogonality constraints on active attractors and suppression of non-active attractors are intended to keep the embedding space structured and prevent false speaker activations.

What would settle it

Retrain one of the stronger baselines, such as the Attribute Attractors plus Conformer system or EEND-EDA-deep, on exactly the same training mixtures, the same 50-second random segments, the same MUSAN augmentation, and the same optimizer and schedule as the proposed system, then evaluate both on the same CALLHOME test split. If that baseline also reaches a DER near 4.99%, the claimed advantage of the new architecture and losses would not be supported.

Watch

Extended reading notes

Core claim

The paper claims that its EEND-CD architecture, which stacks five Latte-based conformer decoders with cross-attention to attractors and self-attentive pooling, achieves a DER of 4.99% on the CALLHOME test set, compared with 9.96% for EEND-EDA, 8.50% for EEND-EDA-deep, and 6.98% for the conformer-enhanced Attribute Attractors system, while using far fewer parameters than the larger baselines. The reported error is broken into 3.60% missed speech, 1.12% false alarm, and 0.27% confusion. The key stated reason for the improvement is the A-DPCL auxiliary loss, which constructs label-attractor vectors by multiplying the ground-truth speaker activity matrix with the current attractor matrix, normalizes those vectors, and then applies a pairwise inner-product loss that aligns audio embeddings with attractor directions. An additional orthogonality loss on active attractors and a suppressive binary cross-entropy loss on inactive attractors are said to reduce false activations and cross-speaker confusion.

Load-bearing premise

The headline comparison assumes that the published diarization error rates of the earlier systems were obtained under the same training data and evaluation conditions, but the authors' training setup differs from at least one baseline because they lack Switchboard-2 Phase I and SRE data and they use mixtures with more than two speakers.

Editorial extensions

If this is right

  • If the reported result holds, end-to-end diarization can reach low error rates with a 15.3-million-parameter model, roughly half the parameter count of the strongest comparison system.
  • Reintroducing deep clustering style losses into EEND, after they were absent from recent attractor-based EEND models, appears to reduce false alarms and confusion rather than just missed speech.
  • Reducing the conformer depth from 12 layers to 5 layers, while adding per-layer attractor decoders, improves both efficiency and diarization error, suggesting that deeper stacks are not necessary when attractors are updated throughout the network.
  • The confusion rate of 0.27% indicates that the attractor orthogonality and angle alignment help separate overlapping speakers, a common failure point in diarization.
  • The paper's own analysis notes that missed speech remains similar to prior systems, so the remaining challenge is reducing misses without raising false alarms or confusion.

Reading between the lines

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

  • Inference: The reported margin over published baselines may be partly due to differences in training data, since the authors lack the Switchboard-2 Phase I and NIST SRE data used by earlier work and they train with mixtures of more than two speakers; retraining baselines under identical conditions would be needed to isolate the architectural contribution.
  • Inference: The A-DPCL idea of building label vectors from attractor prototypes could transfer to other permutation-invariant multi-label problems where the number of active classes varies, such as overlapping event detection, although the paper does not test this.
  • Inference: Because the attractors are compact 256-dimensional vectors updated layer by layer, the architecture may lend itself to streaming or online diarization where a running set of speaker prototypes is maintained, but the paper does not evaluate streaming conditions.
  • Inference: The authors' observation that original DPCL diverges once the model learns to detect speech suggests a general tension between clustering objectives and detection objectives; their label-attractor construction may resolve this tension for other speech tasks that combine embedding structure with binary detection.
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.

Referee Report

5 major / 5 minor

Summary. The paper proposes an end-to-end speaker diarization system, EEND-CD, built from a CNN front-end, a stack of Latte-based conformer decoders, transformer-updated attractors, cross-attention to attractors, self-attentive pooling, and two new auxiliary losses (MO-DPCL and A-DPCL) together with an orthogonality-constrained attractor loss and a permutation-invariant BCE loss. The final five-layer system is reported to achieve 4.99% DER on the CALLHOME test set with 15.3M parameters, compared against several published EEND baselines and against two internal ablations. The paper's claims are that this architecture improves speaker separation, reduces false alarms and confusions, and does so with a compact model.

Significance. If the headline result holds under controlled conditions, the paper would make a useful contribution: it demonstrates that a compact conformer-attractor architecture with a DPCL-style angle loss can reach state-of-the-art DER on CALLHOME, and the parameter count is substantially lower than several compared systems. The internal ablation (rows 6-8 of Table 1) and the explicit statement of parameter counts are strengths, and the test-set DER is a measured number rather than a fitted quantity. However, the external comparison in Table 1 is not controlled for training-data differences, and the claimed significance is not supported by statistical tests. The contribution is therefore plausible but needs additional evidence before the central claim can be accepted.

major comments (5)
  1. [Section 5.1 and Table 1] The central claim that the final system 'achieves a DER of 4.99%, significantly outperforming existing methods' is not supported by a controlled comparison. Section 4.1 states that the authors lack Switchboard-2 Phase I and NIST SRE data, and that unlike [6] they generate mixtures with more than two speakers. Rows 1-4 of Table 1 are published numbers from models trained under different data conditions, so the reported improvement over these rows could be partly or wholly due to training-data composition rather than to the proposed architecture. Please retrain the closest baselines (at least [6] and [7]) under the same data and augmentation protocol, or explicitly restrict the headline claim to the internal ablation and present the external rows only as context.
  2. [Section 5.1 and Section 5.2] The statement in Section 5.1 that the final system 'achieves the lowest confusion rate (0.27%)' is contradicted by Table 1: row 6, the authors' own original model, has a confusion rate of 0.22%, and Section 5.2 acknowledges that the final model has a small increase in confusion. Please correct the text or clarify that the comparison is only against rows 1-4.
  3. [Sections 3.6, 3.7, and 4.6] The combined training objective is not specified precisely. The MSE-based orthogonality loss, the suppressive BCE term, and the operation of 'pushing' non-active attractors to zero are described only verbally, with no equations; the relative weighting of the BCE and DPCL losses, the value of epsilon in Eq. (6), and the rule for deciding which attractors are active versus suppressed are not reported. Without these details the method is not fully reproducible and the contribution of each loss term cannot be assessed.
  4. [Section 5.2 and Table 1] The ablation from row 7 to row 8 changes two factors simultaneously: it introduces the SAP mechanism and replaces MO-DPCL with A-DPCL. Since A-DPCL is stated as one of the paper's key contributions, the individual effect of A-DPCL is not isolated. Please add separate ablation rows so that the contribution of SAP and A-DPCL can be distinguished.
  5. [Section 5.1] The word 'significantly' is used in the sense of a statistical claim, but no error bars, multiple training seeds, or significance tests are reported. The DER difference between the final system and the closest baseline could be within run-to-run variance. Please report results over multiple seeds with mean and variance, and apply a matched-pair significance test, or use weaker wording such as 'numerically lower'.
minor comments (5)
  1. [Section 4.2] The text says '15 consecutive frames ... with a 10 ms hop, resulting in 345-dimensional feature vectors that represent 100 ms of audio.' Fifteen 10 ms frames span 150 ms, not 100 ms; please reconcile this with the window/hop description in Section 3.1.
  2. [Table 1] The multi-row table header is difficult to parse. Please label the subcolumns explicitly (e.g., DER-MS, DER-FA, DER-CF, SAD-MS, SAD-FA) in a single clear header row.
  3. [Section 3.6] Equation (5) uses inner products ⟨l_i, l_j⟩ and ⟨x_i, x_j⟩, but it is not stated explicitly that x_i and x_j are rows of the L2-normalized embedding matrix X. Please clarify the notation.
  4. [Section 3.7] Please clarify how the last layer's attractors are 'split' into a vector a_s and a bias b_s, and whether all S=8 attractors are used during inference or only those assigned to active speakers.
  5. [Section 5.3 and Section 3.5] There are small typos: 'the the use of independent speaker attractors' in Section 5.3 and 'Each attractor as thus evolves' in Section 3.5 should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: held-out DER evaluation, training-objective losses, and non-load-bearing self-citation.

full rationale

The manuscript's central claim (Section 5.1, DER 4.99% on the CALLHOME test set) is a held-out evaluation of a trained model, not a quantity defined in terms of a fitted parameter. The DPCL, MO-DPCL, and A-DPCL losses (Section 3.6) are training-time regularizers; their label-attractor targets are constructed from ground-truth labels and the current attractors, but the reported DER is computed from predictions on the test set and is not algebraically equal to any of these loss terms. Nothing in the paper defines a predicted quantity as the value of a fit: all model parameters (conformer blocks, attractors, biases) are learned by gradient descent, and the final metric is measured, not derived from the training objective. The only self-citation is [7], the authors' prior Attribute Attractors paper, which supplies two rows in Table 1; those rows are used as published baseline values, not as a theorem that forces the present architecture, and the final system's 4.99% result is independently measured against the same CALLHOME test set. Potential concerns about training-data comparability across Table 1 baselines (Section 4.1 vs. [6]) are correctness or attribution risks, not circularity: they do not make the predicted DER an input by construction. No circular step can be exhibited; the derivation chain is self-contained in the sense relevant to this review.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The performance claim rests on hand-chosen architectural hyperparameters and on the comparability of the CALLHOME evaluation to published baselines. No new physical or mathematical entities are introduced; the attractors are learned representations.

free parameters (4)
  • number of attractors S = 8
    Set to 8 in Section 4.5 to track concurrent speakers; chosen by hand, not tuned on validation.
  • number of conformer/attractor layers = 5 (reduced from 12)
    Section 5.2 ablates layer count as a design choice to reduce parameters; the reduction is part of the reported architecture, not derived.
  • Latte latent hidden dimension = 128
    Section 4.5 sets latent attention dimension to 128 alongside 256-dimensional embeddings; chosen by hand.
  • window size / hop = 15 frames / 10 ms hop
    Feature stacking parameters from Section 4.2; standard choices, not fitted.
assumptions (3)
  • domain assumption CALLHOME test set is a valid benchmark and the published baseline DERs are comparable despite training data differences
    Table 1 compares DER to [5,6,7] without retraining baselines on the same data; Section 4.1 notes omitted corpora and different mixture construction.
  • domain assumption The 23-dimensional log-Mel features at 10 ms frame rate retain sufficient information for diarization
    Feature choice in Section 4.2 follows prior EEND work but is not validated independently.
  • standard math Standard backpropagation and AdamW optimization will converge for the combined loss
    The paper relies on standard training machinery without formal convergence guarantees.

how reviews work

0 comments
Cite this review

Pith. "Pith review of End-to-End Diarization utilizing Attractor Deep Clustering." pith.science (2026). https://pith.science/paper/SVLDCLJV

@misc{pith2026250611090,
  author       = {Pith},
  title        = {Pith review of: End-to-End Diarization utilizing Attractor Deep Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SVLDCLJV}},
  note         = {Machine review of arXiv:2506.11090}
}
read the original abstract

Speaker diarization remains challenging due to the need for structured speaker representations, efficient modeling, and robustness to varying conditions. We propose a performant, compact diarization framework that integrates conformer decoders, transformer-updated attractors, and a deep clustering style angle loss. Our approach refines speaker representations with an enhanced conformer structure, incorporating cross-attention to attractors and an additional convolution module. To enforce structured embeddings, we extend deep clustering by constructing label-attractor vectors, aligning their directional structure with audio embeddings. We also impose orthogonality constraints on active attractors for better speaker separation while suppressing non-active attractors to prevent false activations. Finally, a permutation invariant training binary cross-entropy loss refines speaker detection. Experiments show that our method achieves low diarization error while maintaining parameter count.

Figures

Figures reproduced from arXiv: 2506.11090 by the authors.

Figure 1
Figure 1. Example four layer EEND-CD architecture. Each block is described in subsections of Section 3. vectors and audio embeddings, leveraging orthogonality con￾straints on active attractors to reduce cross-speaker confusion. Finally, we include a permutation invariant training (PIT) [8] binary cross-entropy (BCE) loss term to refine speaker detec￾tion, effectively suppressing non-active attractors and minimiz￾ing false ala… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 38 canonical work pages

  1. [6]

    Feed-Forward (FF): A positionwise two-layer MLP with a residual connection

    work page

  2. [7]

    This captures global temporal relationships

    Latent Self-Attention (Latte): A multi-head latent-attention mechanism applied to the time dimension [9]. This captures global temporal relationships

    work page

  3. [1]

    End-to-End Diarization utilizing Attractor Deep Clustering

    Introduction Speaker diarization, the task of segmenting an audio stream into homogeneous speaker segments, remains a challenging prob- lem due to the inherent variability in acoustic conditions, the presence of overlapping speech, and the need to robustly handle a wide range of speakers [1, 2]. End-to-end neural diarization systems (EEND) [3] allow for d...

    work page Pith review arXiv 2025

  4. [2]

    Early methods often combined Gaussian Mixture Models (GMM) or i-vector

    Related Work Speaker diarization involves partitioning an audio stream ac- cording to speaker identity, a task that has traditionally relied on statistical modeling and clustering approaches. Early methods often combined Gaussian Mixture Models (GMM) or i-vector

    work page

  5. [3]

    grayscale

    Proposed Method In this section, we introduce our diarization framework, which combines a CNN-based feature extractor, conformer blocks with latent self-attention (Latte), attractor-based cross-attention, depthwise pooled residual connections, and a DPCL-style angle loss for structured embeddings. Figure 1 illustrates the overall architecture. 3.1. Overvi...

    work page

  6. [4]

    Cross-Attention to Attractors : The output of the CNN is used as queries, while the current speaker attractors serve as keys and values (see Section 3.5)

    work page

  7. [5]

    CNN: A second convolution block for further refinement

    work page

  8. [8]

    CNN: A depthwise convolution module that captures local context

    work page

Show all 46 references
  1. [9]

    Latte: Latent attention for linear time transformers,

    R. Dolga, L. Maystre, M. Cobzarenco, and D. Barber, “Latte: Latent attention for linear time transformers,” 2024. [Online]. Available: https://arxiv.org/abs/2402.17512

    2024

  2. [10]

    representations with clustering [1, 11], later transitioning to x-vectors [12] for more robust speaker embeddings. Accord- ing to Park et al., [2], ”many traditional speaker diarization sys- tems, especially clustering-based systems, have only focused on non-overlapping region...

  3. [11]

    Each conformer block outputs Batch × Time × 256, which is then used to compute speaker logits or passed to subsequent layers (Section 3.4)

    Feed-Forward (FF): Closes the block with another FF sub- layer and residual. Each conformer block outputs Batch × Time × 256, which is then used to compute speaker logits or passed to subsequent layers (Section 3.4). This refinement strategy allows complex interactions between...

  4. [12]

    Stack these embeddings along a depth axis: H(d−1) stack ∈ RBatch×Time×(d)×256. (2)

  5. [13]

    Pool across this depth dimension using self-attentive pooling, yielding P(d−1) ∈ RBatch×Time×256

  6. [14]

    Residually add P to the current input (the most recent audio embeddings): X(d−1) = X(d−1) + P(d−1). (3)

  7. [15]

    This ensures each block sees a global summary of all prior representations, acting as a global residual connection, and smooths training convergence

    Feed X(d−1) into the conformer block, d. This ensures each block sees a global summary of all prior representations, acting as a global residual connection, and smooths training convergence. 3.5. Attractor Updates via Transformer Decoder We initialize a set of attractors A ∈ R...

  8. [16]

    Data Preparation Following the example of prior EEND studies [6], we prepare data from the CALLHOME and Switchboard-2 (Phase II, III) corpora

    Experiments 4.1. Data Preparation Following the example of prior EEND studies [6], we prepare data from the CALLHOME and Switchboard-2 (Phase II, III) corpora. Unlike some earlier works, we do not have access to the Switchboard-2 (Phase I) or the NIST Speaker Recogni- tion Eva...

    2000

  9. [17]

    EEND-EDA [5] 6.4M 6.4M 9.96 5.40 1.36 2.81 3.85 0.87

  10. [18]

    EEND-EDA-deep [6] 29.3M 17.0M 8.50 4.43 1.31 2.76 3.15 0.85

  11. [19]

    and offered alternative loss formulations

  12. [20]

    EEND with Attribute Attractors [7] (12 layers) 33.7M 33.7M 7.87 4.18 1.42 2.27 3.28 0.84

  13. [21]

    EEND with Attribute Attractors + Conformer [7] (12 layers) 35.3M 35.3M 6.98 3.63 1.95 1.41 3.68 0.82

  14. [22]

    7 + SAP + A-DPCL (final system, 5 layers) 15.3M 15.3M 4.99 3.60 1.12 0.27 2.45 0.47 Ablation study on changes required to get to final system

  15. [23]

    EEND with Conformer Decoder (12 layers) 22.2M 22.2M 5.78 4.11 1.45 0.22 3.05 0.51 + last layer Attractor Decoder + MO-DPCL

  16. [24]

    6 + per layer Attractor Decoder (reduce to 5 layers) 14.8M 14.8M 5.60 3.52 1.72 0.36 2.49 0.68 Table 1: Diarization error rates (DER) for each model on the CALLHOME test set and component results. 4.5. Model Configuration For all experiments, the CNN front-end downsamples the ...

  17. [25]

    Results and Discussion 5.1. Evaluation on CALLHOME Table 1 presents the diarization performance of our proposed system and several baseline models on the CALLHOME test set, measured in terms of diarization error rate (DER), missed speech (MS), false alarm (FA), and confusion (...

  18. [26]

    Speaker diarization: A review of recent research,

    X. Anguera, S. Bozonnet, N. W. D. Evans, C. Fredouille, D. Ma- son, and E. S. Chng, “Speaker diarization: A review of recent research,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 20, no. 2, pp. 356–370, 2012

    2012

  19. [27]

    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. [Online]. Available: https://www.sciencedirect. com/science/article/pii...

    2022

  20. [28]

    End-to-end neural speaker diarization with permutation-free ob- jectives,

    Y . Fujita, N. Kanda, S. Horiguchi, S. Watanabe, and T. Nakatani, “End-to-end neural speaker diarization with permutation-free ob- jectives,” in Interspeech 2019, 2019, pp. 4300–4304

    2019

  21. [29]

    Deep clustering: Discriminative embeddings for segmentation and sep- aration,

    J. R. Hershey, Z. Chen, J. L. Roux, and S. Watanabe, “Deep clustering: Discriminative embeddings for segmentation and sep- aration,” in 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016, pp. 31–35

    2016

  22. [30]

    End-to-end speaker diarization for an unknown number of speak- ers with encoder-decoder based attractors,

    S. Horiguchi, Y . Fujita, N. Kanda, S. Watanabe, and T. Nakatani, “End-to-end speaker diarization for an unknown number of speak- ers with encoder-decoder based attractors,” in Interspeech 2020, 2020, pp. 269–273

    2020

  23. [31]

    Neu- ral diarization with non-autoregressive intermediate attractors,

    Y . Fujita, T. Komatsu, R. Scheibler, Y . Kida, and T. Ogawa, “Neu- ral diarization with non-autoregressive intermediate attractors,” in 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  24. [32]

    Improving neural diarization through speaker attribute attractors and local dependency modeling,

    D. Palzer, M. Maciejewski, and E. Fosler-Lussier, “Improving neural diarization through speaker attribute attractors and local dependency modeling,” in 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 11 911–11 915

    2024

  25. [33]

    Permutation invariant training of deep models for speaker-independent multi-talker speech separation,

    D. Yu, M. Kolbæk, Z.-H. Tan, and J. Jensen, “Permutation invariant training of deep models for speaker-independent multi-talker speech separation,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE Press, 2017, p. 241–245. [Online]...

    2017

  26. [34]

    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 Trans- actions on Audio, Speech, and Language Processing , vol. 19, no. 4, pp. 788–798, 2011

    2011

  27. [35]

    Speaker diarization with PLDA i-vector scoring and unsupervised calibration,

    G. Sell and D. Garcia-Romero, “Speaker diarization with PLDA i-vector scoring and unsupervised calibration,” in2014 IEEE Spo- ken Language Technology Workshop (SLT) . IEEE, 2014, pp. 413–417

    2014

  28. [36]

    X-vectors: Robust DNN embeddings for speaker recog- nition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khu- danpur, “X-vectors: Robust DNN embeddings for speaker recog- nition,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5329– 5333

    2018

  29. [37]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,” in Interspeech 2020, 2020, pp. 5036–5040

    2020

  30. [38]

    Robust end-to-end speaker di- arization with conformer and additive margin penalty,

    T.-Y . Leung and L. Samarakoon, “Robust end-to-end speaker di- arization with conformer and additive margin penalty,” in Inter- speech 2021, 2021, pp. 3575–3579

    2021

  31. [39]

    End-to-end neural diarization: From transformer to conformer,

    Y . C. Liu, E. Han, C. Lee, and A. Stolcke, “End-to-end neural diarization: From transformer to conformer,” inInterspeech 2021, 2021, pp. 3081–3085

    2021

  32. [40]

    Linformer: Self-attention with linear complexity,

    S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma, “Linformer: Self-attention with linear complexity,” 2020. [Online]. Available: https://arxiv.org/abs/2006.04768

    2020 arXiv

  33. [41]

    Deepseek-V3 technical report,

    DeepSeek-AI and E. A. Liu, “Deepseek-V3 technical report,”

  34. [43]

    Single-channel multi-speaker separation using deep clustering,

    Y . Isik, J. L. Roux, Z. Chen, S. Watanabe, and J. R. Hershey, “Single-channel multi-speaker separation using deep clustering,” in Interspeech 2016, 2016, pp. 545–549

    2016

  35. [44]

    Alternative objective functions for deep clustering,

    Z.-Q. Wang, J. Le Roux, and J. R. Hershey, “Alternative objective functions for deep clustering,” in 2018 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2018, pp. 686–690

    2018

  36. [45]

    MUSAN: A Music, Speech, and Noise Corpus,

    D. Snyder, G. Chen, and D. Povey, “MUSAN: A Music, Speech, and Noise Corpus,” 2015, arXiv:1510.08484v1

    2015 arXiv

  37. [46]

    From simu- lated mixtures to simulated conversations as training data for end- to-end neural diarization,

    F. Landini, A. Lozano-Diez, M. Diez, and L. Burget, “From simu- lated mixtures to simulated conversations as training data for end- to-end neural diarization,” in Interspeech 2022, 2022, pp. 5095– 5099

    2022

  38. [2024]

    Available: https://arxiv.org/abs/2412.19437

    [Online]. Available: https://arxiv.org/abs/2412.19437

    arXiv

Pith tools

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