Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Improving Neural Diarization through Speaker Attribute Attractors and Local Dependency Modeling

T0 review · 4 major / 6 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 The paper claims that replacing direct intermediate speaker predictions with fixed-size attribute attractors and using a conformer encoder improves diarization, reaching 6.98% DER on CALLHOME versus 8.50% for its re-implemented…

desk verdict A plausible new conditioning mechanism for EEND-EDA, with an honest but under-powered comparison; worth a careful referee, not a desk reject. read the letter →

arxiv 2506.05593 v1 pith:KUAFP7ML submitted 2025-06-05 cs.SD cs.AIeess.AS

David Palzer , Matthew Maciejewski , Eric Fosler-Lussier This is my paper
classification cs.SDcs.AIeess.AS
keywords speakerdiarizationattributeattractorsEEND-EDAconformerintermediateend-to-end
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 speaker diarization model can be improved by changing what intermediate layers are conditioned on: instead of feeding predicted speaker activities back into the network, the model learns a fixed set of 'attribute attractors' that over-segment the speaker embedding space, and later layers attend to these attributes. The paper also replaces the transformer encoder with a conformer to add local temporal context. On the CALLHOME test set, the combined system reaches a diarization error rate of 6.98%, compared with 8.50% for the re-implemented EEND-EDA-deep baseline and 9.96% for the re-implemented EEND-EDA baseline. A sympathetic reading is that attribute-attractor conditioning plus conformer encoders is a practically useful upgrade for neural diarization, with the largest gains in missed-speech detection.

What carries the argument

The load-bearing object is the layer-wise attribute-attractor set $A^a_l \in \mathbb{R}^{N \times D}$, where $N$ is a fixed hyper-parameter larger than the number of speakers $S$. It is a non-autoregressive, over-complete segmentation of the speaker embedding space: instead of forcing each layer to solve the full diarization problem, the model learns $N$ attribute vectors that later layers can selectively attend to. Conditioning is done by multi-head cross-attention, $\hat{E}_l = E_l + \mathrm{MHA}(E_l, A^a_l)$, and a second EDA stage consumes $A^a_l$ to emit speaker attractors $A^s_l$ used only for auxiliary training losses. The second mechanism is the conformer encoder, which adds convolution-augmented local dependencies to the transformer's global attention.

What would settle it

Train the re-implemented baselines with the full original data sources (Switchboard-2 Phase I and NIST SRE) and evaluate all systems on CALLHOME over multiple random seeds; if the final system's DER advantage over the EEND-EDA-deep baseline collapses to roughly zero, the architectural benefit is not supported.

Watch

Extended reading notes

Core claim

The central discovery claimed is that the two-stage EDA design--first producing many attribute attractors, then deriving speaker attractors from them--conditions deeper layers better than direct speaker-activity predictions do. The paper reports that this change, along with multi-head cross-attention conditioning and a conformer backbone, lowers DER on CALLHOME from 8.50% (EEND-EDA-deep re-implementation) to 6.98%, and that the improvement appears layer by layer. The authors attribute the remaining gain to the conformer's local temporal modeling, which primarily reduces missed speech but raises false alarms slightly.

Load-bearing premise

The reported gain rests on the fidelity of the re-implemented baselines, which were trained without Switchboard-2 Phase I and NIST Speaker Recognition Evaluation data that the original systems had, so part of the gap could come from weaker baselines rather than the new architecture.

Editorial extensions

If this is right

  • The principle that intermediate layers should be conditioned on an over-complete attribute representation rather than on direct speaker predictions transfers to other end-to-end diarization architectures.
  • If the reported numbers hold, the 1.52 percent absolute DER gain over the EEND-EDA-deep re-implementation is attributable to the proposed architecture, not to extra training data.
  • Since the final system's layer-wise DER improves with depth, attribute-attractor conditioning provides useful learning signal throughout the stack, not just at the output.
  • The conformer's local temporal modeling is what pushes missed speech down from 4.43% to 3.63%, which suggests that local context is a productive axis for further diarization work.

Reading between the lines

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

  • This suggests the attribute-attractor conditioning could be applied in streaming diarization, where a fixed-size intermediate representation would avoid depending on the full recording.
  • The same cross-attention conditioning may also reduce speaker confusion in overlapping speech, but the paper does not test that separation directly.
  • One testable extension would be to vary the attribute count $N$; the paper fixes it at 256, and the sensitivity of the gain to that choice is unknown.
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

4 major / 6 minor

Summary. The paper proposes two architectural extensions to end-to-end neural diarization with encoder-decoder attractors. First, it replaces direct intermediate speaker-attractor conditioning with 'speaker attribute attractors': a fixed-size set of N=256 non-autoregressive attribute embeddings produced by an EDA module and injected into deeper encoder layers through multi-head cross-attention, while intermediate speaker attractors are still produced from the attribute attractors for auxiliary losses. Second, it swaps the transformer backbone for a conformer encoder to add local temporal dependencies. The models are trained on simulated mixtures from Switchboard-2 (phases II and III) and CALLHOME, and evaluated on the CALLHOME test set. The final system achieves 6.98% DER versus 8.50% for the re-implemented EEND-EDA-deep baseline and 9.96% for the re-implemented EEND-EDA baseline. The paper also reports an ablation chain from the EEND-EDA-deep baseline to the final model and a layer-wise analysis showing monotonic improvement in intermediate layer predictions.

Significance. If the reported gains are robust, the attribute-attractor mechanism is a useful conceptual contribution: it decouples intermediate conditioning from the difficult task of directly predicting speaker labels at every layer, and the switch to conformer encoders is a sensible way to add local inductive bias. The paper is transparent about its training configuration, reports parameter counts, and selects models by validation loss rather than test performance, which avoids the most common form of test-set leakage. The layer-wise results in Table 2 are internally consistent with the design motivation. However, the significance is conditional on two unsecured supports: the baselines are re-implementations on a reduced training corpus, and all headline numbers are single runs with no variance estimates. The architectural idea is promising, but the magnitude of the improvement over the published baselines is not yet established.

major comments (4)
  1. [Section 4.1, Table 1] The headline comparison in Table 1 is against re-implemented baselines trained on a reduced corpus. Section 4.1 states: 'we lack some of the data-sources that they have, namely Switchboard-2 Phase I and NIST Speaker Recognition Evaluation.' Since the original EEND-EDA and EEND-EDA-deep studies used these sources, the reported gap between the proposed systems and the re-implemented baselines may partly reflect a weaker baseline rather than the architectural improvements. The authors should either train all compared systems on identical data that includes the missing sources, report the published results from the original papers under a comparable evaluation, or quantify the effect of adding those data sources to the re-implemented baselines.
  2. [Table 1] Every row of Table 1 appears to be a single run on the CALLHOME test set, with no confidence intervals, standard deviations, or significance tests. Diarization training is sensitive to initialization and data sampling; the differences between adjacent ablation rows (e.g., 8.50% vs. 7.87% vs. 6.98% DER) are small enough that they could fall within run-to-run variability. Please report results over multiple seeds with means and standard deviations, and use paired comparisons or a significance test where appropriate. This is needed to support the central quantitative claim.
  3. [Section 4.1, model configuration 2] The paper also departs from the original EEND-EDA-deep protocol by not limiting training to two-speaker mixtures, stating 'we do not limit ourselves to 2-speaker mixtures as [20] does due to our ability to handle arbitrary numbers of speakers.' It also omits reverberation augmentation. This means the re-implemented EEND-EDA-deep is not a faithful reproduction of the published model, and the comparison may not represent the actual state of the art the paper claims to improve upon. Please either match the original training protocol where feasible or explicitly discuss how these differences affect the comparison.
  4. [Section 4.4, rows 3 and 4 of Table 1] The paper attributes the final DER improvement to the conformer's local temporal modeling, but the comparison between rows 3 and 4 includes all prior ablation changes in addition to the backbone swap. There is no ablation that applies the conformer to the EEND-EDA-deep baseline without attribute attractors, so the specific contribution of local dependency modeling is not isolated. To support the claim about the conformer, please add an ablation that changes only the backbone on an otherwise identical model, or soften the attribution.
minor comments (6)
  1. [Introduction, Section 3.3] There are several typos and minor wording issues: 'introducespeaker' in the Introduction, 'auxilary' in Section 3.3, and 'autogresssive' in Section 3.3. Please proofread carefully.
  2. [Equation (9)] The multi-head cross-attention conditioning in Eq. (9) is not fully specified: please state the number of attention heads, the query/key/value projections, and whether the attribute attractors are used as keys and values with frame embeddings as queries.
  3. [Figure 1] The text references Figure 1, but the figure is not included in the submitted text. Please ensure the final version contains the figure and that the caption explains the asterisk notation for layers removed at inference.
  4. [Section 4.2, Table 1] The 'Total' versus 'Free' parameter columns are not defined. Clarify what 'Free' means and why the EEND-EDA-deep baseline has more total than free parameters, while the other models do not.
  5. [Section 3.2 and Eq. (4)] The notation in Eq. (4) uses \hat{E}_{l-1} without defining how \hat{E}_l relates to E_l and A_l; a short definition would improve readability.
  6. [Section 4.4, Table 2] The layer-wise DER in Table 2 is said to contrast with the original EEND-EDA-deep study, but the authors do not explain why the monotonic improvement occurs or what it implies for the training objective. One or two sentences of interpretation would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical evaluations against held-out CALLHOME labels, with no fitted quantity renamed as a prediction and no load-bearing self-citation chain.

full rationale

The paper's central claim is that attribute-attractor conditioning and a conformer backbone reduce diarization error rate relative to re-implemented EEND-EDA baselines. Every reported number is produced by training a neural network on a training set, selecting the best model by validation loss, and evaluating once on the CALLHOME test set. The model weights are the only fitted quantities, and they are fitted to training labels, not to the test labels being reported; no parameter of the model is fit on the CALLHOME test set and then presented as a prediction. The hyperparameter N=256 for attribute attractors is a fixed architectural choice stated in Section 4.2, and the paper does not tune it on the test set or derive it from the target result. The ablation chain in Table 1 measures incremental architecture changes under the same training and evaluation protocol, so the comparisons are internally coherent. The paper's caveat in Section 4.1 that the baselines lack Switchboard-2 Phase I and NIST SRE data concerns whether the comparison is fair or whether the reimplemented baselines are as strong as the published systems; this is an external-validity and baseline-fidelity concern, not a circularity concern. Similarly, the absence of error bars and significance tests affects robustness of the claimed gain, but does not make any derivation reduce to its own inputs. The architectural components (attribute attractors, cross-attention conditioning, transformer-based EDA, conformer) are described by their own equations in Section 3 and are not justified by citing the authors' own prior work as the sole evidence for a contested premise. I find no step where an output is defined in terms of the target quantity, no fitted input is relabeled as a prediction, and no load-bearing result is imported from a self-citation. The finding is therefore a clean non-finding: score 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a supervised training setup with no free parameters fitted to the test set; the sole hand-chosen hyperparameter is N=256 attribute attractors. No new physical entities are introduced; 'attribute attractors' are latent vectors inside the trained model, and the paper provides no out-of-model evidence for them as a separate construct.

free parameters (1)
  • Number of attribute attractors N = 256
    Chosen by hand as a hyperparameter; the paper notes N is larger than the number of speakers but gives no tuning analysis for sensitivity.
assumptions (4)
  • standard math Neural network training via backpropagation and AdamW converges to a good solution
    Standard assumption in empirical ML papers; no formal convergence guarantee is provided.
  • domain assumption CALLHOME test labels are correct and the DER metric is a valid measure of diarization quality
    The paper uses the standard NIST-style metric without further justification or manual inspection.
  • domain assumption Simulated mixtures from Switchboard-2 and CALLHOME training set are representative of the test distribution
    This is the training setup described in Section 4.1; if the simulated mixtures poorly match real CALLHOME conversations, the results may not generalize.
  • domain assumption The re-implemented baselines match the original implementations despite the missing training data sources
    Section 4.1 states the re-implementations match the original, but also that they lack Switchboard-2 Phase I and NIST SRE data; this is load-bearing for the comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Neural Diarization through Speaker Attribute Attractors and Local Dependency Modeling." pith.science (2026). https://pith.science/paper/KUAFP7ML

@misc{pith2026250605593,
  author       = {Pith},
  title        = {Pith review of: Improving Neural Diarization through Speaker Attribute Attractors and Local Dependency Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KUAFP7ML}},
  note         = {Machine review of arXiv:2506.05593}
}
read the original abstract

In recent years, end-to-end approaches have made notable progress in addressing the challenge of speaker diarization, which involves segmenting and identifying speakers in multi-talker recordings. One such approach, Encoder-Decoder Attractors (EDA), has been proposed to handle variable speaker counts as well as better guide the network during training. In this study, we extend the attractor paradigm by moving beyond direct speaker modeling and instead focus on representing more detailed `speaker attributes' through a multi-stage process of intermediate representations. Additionally, we enhance the architecture by replacing transformers with conformers, a convolution-augmented transformer, to model local dependencies. Experiments demonstrate improved diarization performance on the CALLHOME dataset.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 36 canonical work pages

  1. [20]

    End-to-end neural speaker di- arization with self-attention,

    Y . Fujita, N. Kanda, S. Horiguchi, Y . Xue, K. Naga- matsu, and S. Watanabe, “End-to-end neural speaker di- arization with self-attention,”ASRU, p. 296–303, 2019

    work page 2019

  2. [1]

    INTRODUCTION Speaker diarization is the activity of labeling an audio record- ing with continuous segments by speaker identity. Diariza- tion has a large range of applications, such as pre-processing audio for downstream tasks like source separation, informa- tion retrieval for pre-recorded audio, speaker-turn analysis for call recordings, and processing ...

    work page

  3. [2]

    Improving Neural Diarization through Speaker Attribute Attractors and Local Dependency Modeling

    RELATED WORK Traditional diarization solutions [9, 10] utilize clustering on speaker embeddings such as i-vectors [11], d-vectors [12], and x-vectors [13]. These speaker time embeddings are clus- tered using traditional ML algorithms, such as GMMs [9], or hierarchical clustering [14]. Neural diarization techniques like EEND [5] employ an end-to-end neural...

    work page Pith review arXiv 2025

  4. [3]

    METHODS This section describes the original EEND-EDA [21] and EEND-EDA-deep [20] architectures followed by our pro- posed improvements to both models. 3.1. EEND-EDA EEND-EDA modifies the original EEND network to use a stack ofLtransformers instead of LSTMs and to allow for an arbitrary number of speakers, which is not known a priori. El = Encoder(El−1) (1...

    work page

  5. [4]

    EXPERIMENTS 4.1. Data In following the example of the line of EEND studies, we prepare the CALLHOME, and Switchboard-2 (phase II, III) the same way, however we lack some of the data-sources that they have, namely Switchboard-2 Phase I and NIST Speaker Recognition Evaluation. Additionally we do not limit our- selves to 2-speaker mixtures as [20] does due t...

    work page 2000

  6. [5]

    2 + Non-shared EDA 29.3M 29.3M 8.79 4.98 1.46 2.35 3.73 0.85

    work page

  7. [6]

    EEND-EDA [21] 6.4M 6.4M 9.96 5.40 1.36 2.81 3.85 0.87

    work page

  8. [7]

    EEND-EDA-deep [20] 29.3M 17.0M 8.50 4.43 1.31 2.76 3.15 0.85

    work page

Show all 37 references
  1. [8]

    EEND with Attribute Attractors 33.7M 33.7M 7.87 4.18 1.42 2.27 3.28 0.84

  2. [9]

    3 + Conformer (Final system) 35.3M 35.3M 6.98 3.63 1.95 1.41 3.68 0.82 Ablation study on changes required to get to EEND with Attribute Attractors

  3. [10]

    Unsu- pervised methods for speaker diarization: An integrated and iterative approach,

    S.H. Shum, N. Dehak, R. Dehak, and J.R. Glass, “Unsu- pervised methods for speaker diarization: An integrated and iterative approach,”IEEE TASLP, vol. 21, no. 10, pp. 2015–2028,, 2013

    2015

  4. [11]

    5 + Cross Attention 32.1M 32.1M 8.48 4.52 1.30 2.66 3.23 0.86

  5. [12]

    6 + TransformerEDA 32.9M 32.9M 7.98 4.22 1.43 2.33 3.36 0.85 Table 1. Diarization error rates (DER) for each model on the CALLHOME test set and component results Layer DER (%) SAD (%) DER MS FA CF MS FA 1 26.07 10.86 8.40 6.81 8.34 0.81 2 21.13 11.39 4.29 5.45 9.28 1.46 3 15.8...

  6. [13]

    Additionally we updated parts of the model with modern attentive layers as well as introduced temporal inductive bias

    CONCLUSION We created an end-to-end neural diarization architecture that uses attribute attractors to condition deeper layer of the net- work on our speakers. Additionally we updated parts of the model with modern attentive layers as well as introduced temporal inductive bias....

  7. [14]

    An overview of au- tomatic speaker diarization systems,

    S.E. Tranter and D.A. Reynolds, “An overview of au- tomatic speaker diarization systems,”IEEE Trans. on ASLP, vol. 14, no. 5, pp. 1557–1565,, 2006

    2006

  8. [15]

    Overlap in meetings: Asr effects and analysis by dialog factors, speakers, and col- lection site,

    ¨O. C ¸ etin and E. Shriberg, “Overlap in meetings: Asr effects and analysis by dialog factors, speakers, and col- lection site,” inProc. MLMI, 2006, p. 212–224

    2006

  9. [16]

    Guided source separation meets a strong asr backend: Hitachi/paderborn university joint investigation for din- ner party scenario,

    N. Kanda, C. Boeddeker, J. Heitkaemper, Y . Fujita, S. Horiguchi, K. Nagamatsu, and R. Haeb-Umbach, “Guided source separation meets a strong asr backend: Hitachi/paderborn university joint investigation for din- ner party scenario,”INTERSPEECH, p. 1248–1252, 2019

    2019

  10. [17]

    Hi- tachi/jhu chime-5 system: Advances in speech recog- nition for everyday home environments using multiple microphone arrays,

    N. Kanda, R. Ikeshita, S. Horiguchi, Y . Fujita, K. Naga- matsu, X. Wang, V . Manohar, N.E.Yalta Soplin, M. Ma- ciejewski, S.J. Chen, A.S. Subramanian, R. Li, Z. Wang, J. Naradowsky, L.P. Garcia-Perera, and G. Sell, “Hi- tachi/jhu chime-5 system: Advances in speech recog- niti...

    2018

  11. [18]

    End-to-end neural speaker diarization with permutation-free objectives,

    Y . Fujita, N. Kanda, S. Horiguchi, K. Nagamatsu, and S. Watanabe, “End-to-end neural speaker diarization with permutation-free objectives,”INTERSPEECH, p. 4300–4304, 2019

    2019

  12. [19]

    Permuta- tion invariant training of deep models for speaker inde- pendent multi-talker speech separation,

    D. Yu, M. Kolbæk, Z.-H. Tan, and J. Jensen, “Permuta- tion invariant training of deep models for speaker inde- pendent multi-talker speech separation,” inICASSP, p. 241–245. 2017

    2017

  13. [21]

    This PyTorch implementation matches the original

    to train on publicly available dataset splits. This PyTorch implementation matches the original. EEND-EDA-deep with Intermediate Attractors (2):This model reimplements [20] using intermediate attractors and in- termediate auxiliary losses on all layers. Non-Shared EDA (5):This...

  14. [22]

    Encoder-decoder based attractors for end-to-end neural diarization,

    Shota Horiguchi, Yusuke Fujita, Shinji Watanabe, Yawen Xue, and Paola Garcia, “Encoder-decoder based attractors for end-to-end neural diarization,”IEEE/ACM Transactions on Audio, Speech, and Language Process- ing, vol. 30, pp. 1493–1507, 2022

    2022

  15. [23]

    Lium spkdiarization: An open source toolkit for diarization,

    S. Meignier, “Lium spkdiarization: An open source toolkit for diarization,” inCMU SPUD Workshop. 2010

    2010

  16. [24]

    Front-end factor analysis for speaker ver- ification,

    N. Dehak, P.J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front-end factor analysis for speaker ver- ification,”IEEE Trans. on ASLP, vol. 19, no. 4, pp. 788–798,, 2011

    2011

  17. [25]

    Speaker diarization with LSTM,

    Q. Wang, C. Downey, L. Wan, P.Andrew Mansfield, and I.Lopez Moreno, “Speaker diarization with LSTM,” in ICASSP, p. 5239–5243. 2018

    2018

  18. [26]

    Speaker diarization using deep neu- ral network embeddings,

    D. Garcia-Romero, D. Snyder, G. Sell, D. Povey, and A. McCree, “Speaker diarization using deep neu- ral network embeddings,” inProc. ICASSP, 2017, p. 4930–4934

    2017

  19. [27]

    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,” SLT, p. 413–417, 2014

    2014

  20. [28]

    Listening to each speaker one by one with recurrent selective hearing networks,

    K. Kinoshita, L. Drude, M. Delcroix, and T. Nakatani, “Listening to each speaker one by one with recurrent selective hearing networks,” inICASSP, p. 5064–5068. 2018

    2018

  21. [29]

    All-neural online source separation, counting, and diarization for meeting analysis,

    T. Neumann, K. Kinoshita, M. Delcroix, S. Araki, T. Nakatani, and R. Haeb-Umback, “All-neural online source separation, counting, and diarization for meeting analysis,” inICASSP, p. 91–95. 2019

    2019

  22. [30]

    Recursive speech separation for unknown number of speakers,

    N. Takahashi, S. Parthasaarathy, N. Goswami, and Y . Mitsufuji, “Recursive speech separation for unknown number of speakers,”INTERSPEECH, p. 1348–1352, 2019

    2019

  23. [31]

    Deep attractor network for single-microphone speaker separation,

    Z. Chen, Y . Luo, and N. Mesgarani, “Deep attractor network for single-microphone speaker separation,” in ICASSP, p. 246–250. 2017

    2017

  24. [32]

    Speaker- independent speech separation with deep attractor net- work,

    Y . Luo, Z. Chen, and N. Mesgarani, “Speaker- independent speech separation with deep attractor net- work,”IEEE/ACM TASLP, vol. 26, no. 4, pp. 787–796,, 2018

    2018

  25. [33]

    Neural diariza- tion with non-autoregressive intermediate attractors,

    Yusuke Fujita, Tatsuya Komatsu, Robin Scheibler, Yusuke Kida, and Tetsuji Ogawa, “Neural diariza- tion with non-autoregressive intermediate attractors,” in ICASSP 2023, 2023, pp. 1–5

    2023

  26. [34]

    End-to-end neural diarization: Reformulating speaker diarization as simple multi-label classification,

    Yusuke Fujita, Shinji Watanabe, Shota Horiguchi, Yawen Xue, and Kenji Nagamatsu, “End-to-end neural diarization: Reformulating speaker diarization as simple multi-label classification,”ArXiv, vol. abs/2003.02966, 2020

    2003 arXiv

  27. [35]

    Conformer: Convolution-augmented transformer for speech recognition,

    Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang, “Conformer: Convolution-augmented transformer for speech recognition,” 2020

    2020

  28. [36]

    MU- SAN: A Music, Speech, and Noise Corpus,

    David Snyder, Guoguo Chen, and Daniel Povey, “MU- SAN: A Music, Speech, and Noise Corpus,” 2015, arXiv:1510.08484v1

    2015 arXiv

  29. [37]

    From simulated mixtures to simulated conversations as training data for end-to-end neural di- arization,

    Federico Landini, Alicia Lozano-Diez, Mireia Diez, and Luk´aˇs Burget, “From simulated mixtures to simulated conversations as training data for end-to-end neural di- arization,” 2022

    2022

Pith tools

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