Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Spatio-spectral diarization of meetings by combining TDOA-based segmentation and speaker embedding-based clustering

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a meeting diarization pipeline which first segments speech by time-difference-of-arrival cues and then clusters speaker embeddings outperforms single-channel neural diarization, works for compact and distributed…

desk verdict A solid spatio-spectral diarization front end that works without in-domain multi-channel training data; the distributed-microphone claim needs a synchronization caveat, but the core result holds. read the letter →

arxiv 2506.16228 v2 pith:4CTTLZFF submitted 2025-06-19 eess.AS cs.SD

Tobias Cord-Landwehr , Tobias Gburrek , Marc Deegen , Reinhold Haeb-Umbach This is my paper · ORCID
classification eess.AScs.SD
keywords speakerdiarizationtimedifferenceofarrivalembeddingsbeamformingoverlappingspeechdistributedmicrophonescompactmicrophonearraymeetingtranscription
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 is trying to establish that meeting diarization — assigning every stretch of a recording to the person who said it — improves when spatial and spectral cues are cascaded rather than used alone. The proposed pipeline estimates time-difference-of-arrival vectors from a multi-channel recording, groups them into speech segments, beamforms each segment, and then clusters the resulting speaker embeddings into identities. The central claim is that this spatio-spectral design gets the overlap-handling advantage of spatial processing and the position-change stability of spectral processing, while requiring neither multi-channel training data nor knowledge of the number or placement of microphones. If correct, multi-channel diarization can be deployed on unknown ad-hoc microphone arrangements and still beat a single-channel spectral system by several percentage points of error rate in both compact-array and distributed settings.

What carries the argument

The load-bearing object is the TDOA vector $\boldsymbol{\tau}=[\tau_{0,1},\tau_{0,2},\ldots,\tau_{C-1,C}]^{\top}$ assembled from pairwise GCC-PhaT delay estimates; a vector survives only if closed-loop delay sums are near zero. Temporally constrained leader-follower clustering of these vectors produces speech segments, and each segment's median TDOA vector acts as a steering prototype. Segment-level beamforming applies masks obtained by comparing spatial covariance matrices to those prototypes, and a cACGMM step refines the masks. The second stage is the speaker embedding extractor (a ResNet34-based d-vector model) followed by HDBSCAN clustering on cosine distances; that spectral stage is what lets reflections and position changes merge into one speaker cluster.

What would settle it

Record a meeting with four ad-hoc devices whose clocks carry a known sampling-rate offset (for example, 100 ppm) and run the pipeline without any synchronization; if the DER stays close to the synchronized case, the synchronization premise is not load-bearing, whereas the spatial segmentation stage is expected to degrade sharply.

Watch

Extended reading notes

Core claim

The central claim is that a meeting diarization system can be built by replacing the learned single-channel segmentation stage of a spectral diarization system with a model-based spatial segmentation stage while keeping embedding-based clustering. The spatial stage estimates multiple TDOA vectors per time frame via GCC-PhaT, retains only vectors whose closed-loop delays sum near zero, and groups them into temporally constrained speech segments. Each segment is beamformed using binary masks derived from spatial-covariance distance to prototype steering vectors, optionally refined by a cACGMM. The spectral stage then extracts a speaker embedding from each beamformed segment and clusters them with HDBSCAN, merging outlier segments and intersecting segments into the same cluster. The paper claims this cascade outperforms single-channel neural diarization on both compact and distributed microphone setups and, unlike a spatial-only system, keeps diarization stable when speakers change positions, because embedding clustering is position-agnostic.

Load-bearing premise

The distributed-microphone results presuppose that all channels are synchronized in sampling-rate and sampling-time offsets, which the paper states as an assumption and does not test.

Editorial extensions

If this is right

  • On the compact-array corpus the pipeline reports an average Diarization Error Rate (DER) of 7.17% and 9.97% in overlap-only regions, while on the distributed corpus it reaches 3.79-3.92%, so overlap handling no longer dominates the error budget.
  • Switching from the compact array to distributed smartphones costs about 1-2 percentage points absolute in DER and concatenated minimum-permutation word error rate (cpWER), with only two delay-threshold parameters changed, so the same trained system transfers across arrangements.
  • In concatenated meetings where several speakers change positions, the spatial-only reference's cpWER jumps from single digits to 24-74%, while the proposed system stays between 3.3% and 6.5%, so the cascade does not lose track when positions change.
  • Because the embedding extractor is trained on single-channel speaker-verification data, the system can be deployed on unseen multi-channel geometries without collecting in-domain multi-channel training data.

Reading between the lines

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

  • A direct implication the paper does not test: before the TDOA stage, real ad-hoc devices would need a sampling-rate and sampling-time synchronization step; until then, the distributed-microphone claim is confined to synchronized recordings.
  • The low-pass loudspeaker incident in one evaluated corpus reveals that the reflection-rejection heuristic implicitly assumes a particular spectral shape for speech; a frequency-band-independent reflection test, or a learned classifier, could remove that failure mode.
  • The same spatial-segmentation-plus-spectral-merging pattern could be applied to other spatial cues, such as directional statistics or position embeddings, whenever position changes or phantom sources are concerns.
  • A testable extension of the paper's semi-static result is to shorten the 1-second temporal window in segment detection so that slowly moving speakers are captured by a chain of short segments, which the paper lists as future work.
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

3 major / 5 minor

Summary. The manuscript proposes a two-stage meeting diarization pipeline. The first stage is model-based and spatial: GCC-PhaT TDOA estimation, temporally constrained leader-follower segment detection over TDOA vectors, segment-level beamforming with binary masks and optional cACGMM refinement. The second stage is data-driven and spectral: d-vector embeddings are extracted from each beamformed segment and clustered with HDBSCAN, with outlier merging and intersection handling. The system is evaluated on LibriCSS and LibriWASN in both compact-array and distributed configurations, and compared against pyannote, a spatial-only pipeline, SSND, and other published systems using DER and cpWER. The authors report large DER reductions over pyannote in the compact setup, and a semi-static experiment shows that the spectral clustering stage can track speakers after position changes where a purely spatial pipeline fails.

Significance. If the reported results hold, the paper makes a useful practical contribution: it demonstrates that a model-based spatial front end can be combined with a pretrained speaker embedding extractor, requiring neither in-domain multi-channel training data nor knowledge of microphone geometry. The compact-setup gains over pyannote are large and consistent across LibriWASN and LibriCSS, and the semi-static experiment gives direct evidence that the spectral clustering stage repairs position-change failures of spatial diarization. The public code and the use of public benchmarks are also strengths. The main caveat is that the distributed-microphone claim is validated only for perfectly synchronized channels, with no experiment addressing asynchronous recording conditions.

major comments (3)
  1. [Section 3.1] The distributed-microphone evaluation relies on the statement 'all channels are assumed to be synchronized both in terms of Sampling Rate Offset (SRO) and Sampling Time Offset (STO).' This assumption is load-bearing: the GCC-PhaT TDOA estimator in Eq. (2), the temporal clustering of TDOA vectors in Section 2.2, and the steering-vector prototypes in Eq. (4) all require stable inter-channel delays. LibriWASN is described in reference [14] as a dataset with asynchronous recording devices, yet the paper reports no synchronization front-end and no asynchronous-channel experiment. The abstract's claim that the system works with distributed microphones is therefore supported only under a strong premise that is not tested. Please either add an experiment with SRO/STO mismatch or qualify the distributed claim accordingly.
  2. [Section 3.3 / Table 2] Table 2 compares the proposed pipeline against pyannote only in the compact microphone setup, yet the abstract states that the proposed system significantly outperforms pyannote both in a compact-array scenario and in a distributed setup. In the distributed condition, Table 1 reports results for the proposed system and spatial-only baseline, but no single-channel spectral baseline such as pyannote is shown. Thus the headline distributed comparison is not directly demonstrated. Please add a distributed pyannote result or restrict the claim to the compact setup.
  3. [Tables 1-3] All DER and cpWER numbers are reported as single aggregated values with no per-meeting breakdown, number of test conversations, confidence intervals, or significance tests. Given the abstract's use of 'significantly outperforms' and the fact that some comparisons (e.g., proposed versus SSND on LibriCSS in Table 2) are close, the paper should report variance across the test meetings or otherwise justify the statistical claim.
minor comments (5)
  1. [Throughout] The dataset name is written inconsistently as 'LibriW ASN' and 'LibriWASN'; please standardize the spelling.
  2. [Table 1] The column headers '0S' and '0L' are never defined; please add a footnote or caption explaining these overlap conditions.
  3. [Section 2.1] The expression 'P (C−1)C/2 different TDOA vectors' is ambiguous; it should be written as P^{C(C-1)/2} or with explicit combinatorial notation.
  4. [Figure 2] The caption does not explain the time and frequency axes or why the highlighted region corresponds to the defective loudspeaker; please clarify the visual evidence.
  5. [Section 3.2] The comparison to references [33] and [34] uses DER OV values from the literature without stating whether the evaluation protocols (collar, overlap definition) are identical; please state protocol equivalence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the spatio-spectral pipeline is an assembled cascade of published components, and the headline claims are decided by external benchmark comparisons, not by construction.

full rationale

The paper's derivation chain is not circular. The TDOA/GCC-PhaT segmentation (Sec. 2.1-2.2), segment-level beamforming (Sec. 2.3), and d-vector/HDBSCAN clustering (Sec. 2.4) are applied as fixed algorithms; no equation defines the measured DER in terms of a fitted parameter, and no fitted quantity is renamed as a prediction. The algorithm-specific thresholds tau_th, Delta_tau_max, and the reflection mask threshold are tuned constants, not outputs of the evaluation, so Tables 1-3 report genuine empirical outcomes. The self-citations to [13] and [26] are load-bearing in the sense that the system reuses prior components, but those components are externally validated, code-released, and are not adjusted to force the DER numbers; therefore they do not make the evaluation circular. The strongest caveat, the SRO/STO synchronization assumption in Sec. 3.1, is a scope limitation that would affect generalization to asynchronous ad-hoc devices, not a circularity: it does not reduce the claimed result to its own input. The comparison with pyannote and SSND on LibriCSS/LibriW ASN is an external falsification test, so the central claim has independent content.

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

The central claims rest on several manually selected thresholds (Section 3.1) and on standard assumptions inherited from the underlying TDOA and beamforming literature. Most consequential are the unstated reflection-rejection threshold that caused the LibriCSS loudspeaker failure, and the explicit synchronization assumption for distributed microphones. No new physical entities are introduced.

free parameters (4)
  • tau_th delay consistency threshold = 1 (compact), 2 (distributed)
    Used to discard TDOA vectors that violate loop consistency; manually set per microphone geometry in Section 3.1; no sensitivity analysis.
  • Delta_tau_max maximum TDOA distance = 1 sample (compact), 0.75 samples (distributed)
    Controls segment clustering in temporally constrained leader-follower clustering; manually set per setup in Section 3.1.
  • Maximum temporal distance between frames in a segment = 1 s
    Segment clustering rule from Section 2.2; chosen system parameter, value not justified experimentally.
  • Reflection rejection mask-activity threshold = not specified
    Segments whose 150-3500 Hz mask activity falls below an unstated threshold are discarded in Section 2.3; this threshold caused the defective-loudspeaker failure on LibriCSS.
assumptions (4)
  • domain assumption W-disjoint orthogonality of speech
    Section 2.3: each tf-bin is assumed dominated by one source or noise, enabling binary masks; approximate and known to degrade in reverberation.
  • domain assumption Distributed microphone channels are synchronized in SRO and STO
    Section 3.1 explicitly assumes synchronization; without it, TDOA estimates fail, limiting the claim of working with distributed devices.
  • domain assumption A speaker position is stationary within each detected segment
    Section 1 states the system requires a source not moving only for a single segment; Section 2.2 groups frames by TDOA similarity over at most 1 s.
  • domain assumption GCC-PhaT maxima correspond to physical sources and loop-consistency removes ghost TDOA vectors
    Section 2.1 relies on this to generate candidate TDOA vectors; in reverberation this is approximate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatio-spectral diarization of meetings by combining TDOA-based segmentation and speaker embedding-based clustering." pith.science (2026). https://pith.science/paper/4CTTLZFF

@misc{pith2026250616228,
  author       = {Pith},
  title        = {Pith review of: Spatio-spectral diarization of meetings by combining TDOA-based segmentation and speaker embedding-based clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4CTTLZFF}},
  note         = {Machine review of arXiv:2506.16228}
}
read the original abstract

We propose a spatio-spectral, combined model-based and data-driven diarization pipeline consisting of TDOA-based segmentation followed by embedding-based clustering. The proposed system requires neither access to multi-channel training data nor prior knowledge about the number or placement of microphones. It works for both a compact microphone array and distributed microphones, with minor adjustments. Due to its superior handling of overlapping speech during segmentation, the proposed pipeline significantly outperforms the single-channel pyannote approach, both in a scenario with a compact microphone array and in a setup with distributed microphones. Additionally, we show that, unlike fully spatial diarization pipelines, the proposed system can correctly track speakers when they change positions.

Figures

Figures reproduced from arXiv: 2506.16228 by the authors.

Figure 1
Figure 1. Illustration of the proposed spatio-spectral diarization pipeline. offers some conclusions and an outlook on future work. 2. Spatio-spectral diarization pipeline A multi-channel recording of a meeting with K speakers can be modeled in the time-frequency domain as the summation yc(t, f) = X l(k) sl(k)(t, f)hc,l(k)(t, f) (1) of delayed, clean speech signals sl(k)(t, f) which are padded to match the length of the conve… view at source ↗
Figure 2
Figure 2. Spectrogram of a LibriCSS segment with a single speech region of the defective loudspeaker (highlighted in red) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages

  1. [14]

    Acoustic beamform- ing for speaker diarization of meetings,

    X. Anguera, C. Wooters, and J. Hernando, “Acoustic beamform- ing for speaker diarization of meetings,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 15, no. 7, pp. 2011–2022, 2007

    work page 2011

  2. [1]

    Spatio-spectral diarization of meetings by combining TDOA-based segmentation and speaker embedding-based clustering

    Introduction Diarization systems assign regions of speech activity to the in- dividual participants of a conversation, thus answering the ques- tion “Who spoke when?”. Essentially, they solve two tasks, seg- mentation and speaker assignment. The first is on identifying regions (segments) of constant speaker activity, while the sec- ond assigns speaker lab...

    work page Pith review arXiv 2025

  3. [2]

    prototype

    Spatio-spectral diarization pipeline A multi-channel recording of a meeting withK speakers can be modeled in the time-frequency domain as the summation yc(t, f) = X l(k) sl(k)(t, f)hc,l(k)(t, f) (1) of delayed, clean speech signals sl(k)(t, f) which are padded to match the length of the conversation. Here, hc,l(k)(t, f) de- notes the acoustic transfer fun...

    work page

  4. [3]

    semi- static

    Experiments 3.1. Experimental Setup For evaluation, the proposed pipeline is applied to the Lib- riCSS [28] and LibriW ASN [14] data sets. LibriCSS consists of re-recordings of simulated LibriSpeech 8-speaker meetings ranging from 0 % to 40 % overlapping speech with a duration of 10 min. LibriW ASN is an additional re-recording of the same synthetic meeti...

    work page

  5. [4]

    The proposed model can be de- ployed in compact and distributed microphone setups without large performance differences and with only minimal parame- ter changes

    Conclusion In this work, we presented an approach that combines spa- tial segmentation with a spectral, embedding-based cluster- ing model for the diarization of meetings without requiring in-domain training data. The proposed model can be de- ployed in compact and distributed microphone setups without large performance differences and with only minimal p...

    work page

  6. [5]

    Acknowledgements Computational Resources were provided by BMBF/NHR/PC2

    work page

  7. [6]

    A DOA based speaker diarization system for real meetings,

    S. Araki, M. Fujimoto, K. Ishizuka, H. Sawada, and S. Makino, “A DOA based speaker diarization system for real meetings,” in Hands-Free Speech Communication and Microphone Arrays (HSCMA), 2008, pp. 29–32

    work page 2008

  8. [7]

    X-vectors: Robust DNN embeddings for speaker recogni- tion,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudan- pur, “X-vectors: Robust DNN embeddings for speaker recogni- tion,” in Proc. IEEE ICASSP, 2018, pp. 5329–5333

    work page 2018

Show all 41 references
  1. [8]

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

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

    2022

  2. [9]

    Target-speaker voice activity detection: A novel approach for multi-speaker diarization in a dinner party sce- nario,

    I. Medennikov, M. Korenevsky, T. Prisyach, Y . Khokhlov, M. Ko- renevskaya et al. , “Target-speaker voice activity detection: A novel approach for multi-speaker diarization in a dinner party sce- nario,” in Proc. ISCA Interspeech, 2020, pp. 274–278

    2020

  3. [10]

    Powerset multi-class cross entropy loss for neural speaker diarization,

    A. Plaquet and H. Bredin, “Powerset multi-class cross entropy loss for neural speaker diarization,” in Proc. ISCA Interspeech , 2023

    2023

  4. [11]

    An overview of automatic speaker diarization systems,

    S. Tranter and D. Reynolds, “An overview of automatic speaker diarization systems,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, no. 5, pp. 1557–1565, 2006

    2006

  5. [12]

    Deep speaker: An end-to-end neural speaker embedding system,

    C. Li, X. Ma, B. Jiang, X. Li, X. Zhang et al., “Deep speaker: An end-to-end neural speaker embedding system,” arXiv preprint arXiv:1705.02304, 2017

    2017 arXiv

  6. [13]

    ECAPA- TDNN: Emphasized Channel Attention, Propagation and Aggre- gation in TDNN Based Speaker Verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA- TDNN: Emphasized Channel Attention, Propagation and Aggre- gation in TDNN Based Speaker Verification,” in Proc. ISCA In- terspeech, 2020, pp. 3830–3834

    2020

  7. [15]

    Probabilistic speaker diarization with bag-of-words representa- tions of speaker angle information,

    K. Ishiguro, T. Yamada, S. Araki, T. Nakatani, and H. Sawada, “Probabilistic speaker diarization with bag-of-words representa- tions of speaker angle information,”IEEE Transactions on Audio, Speech, and Language Processing , vol. 20, no. 2, pp. 447–460, 2011

    2011

  8. [16]

    Modeling audio directional statistics using a probabilistic spatial dictionary for speaker diarization in real meetings,

    M. Fakhry, N. Ito, S. Araki, and T. Nakatani, “Modeling audio directional statistics using a probabilistic spatial dictionary for speaker diarization in real meetings,” inIEEE International Work- shop on Acoustic Signal Enhancement (IWAENC), 2016, pp. 1–5

    2016

  9. [17]

    Spatial-aware speaker diarization for multi-channel multi-party meeting,

    J. Wang, Y . Liu, B. Wang, Y . Zhi, S. Li et al. , “Spatial-aware speaker diarization for multi-channel multi-party meeting,” in Proc. ISCA Interspeech, 2022

    2022

  10. [18]

    Spatial diarization for meeting transcription with ad-hoc acoustic sensor networks,

    T. Gburrek, J. Schmalenstroeer, and R. Haeb-Umbach, “Spatial diarization for meeting transcription with ad-hoc acoustic sensor networks,” in 57th Asilomar Conference on Signals, Systems, and Computers, 2023, pp. 1399–1403

    2023

  11. [19]

    Libri- W ASN: A data set for meeting separation, diarization, and recog- nition with asynchronous recording devices,

    J. Schmalenstroeer, T. Gburrek, and R. Haeb-Umbach, “Libri- W ASN: A data set for meeting separation, diarization, and recog- nition with asynchronous recording devices,” in ITG conference on Speech Communication, Sep 2023

    2023

  12. [20]

    Multi-channel conversational speaker separation via neural diarization,

    H. Taherian and D. Wang, “Multi-channel conversational speaker separation via neural diarization,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  13. [21]

    Incorporat- ing spatial cues in modular speaker diarization for multi-channel multi-party meetings,

    R. Wang, S. Niu, G. Yang, J. Du, S. Qian et al. , “Incorporat- ing spatial cues in modular speaker diarization for multi-channel multi-party meetings,” arXiv preprint arXiv:2409.16803, 2024

    2024 arXiv

  14. [22]

    Multi-channel speaker diarization using spatial features for meetings,

    N. Zheng, N. Li, J. Yu, C. Weng, D. Su et al. , “Multi-channel speaker diarization using spatial features for meetings,” in Proc. IEEE ICASSP, 2022, pp. 7337–7341

    2022

  15. [23]

    Summary of the NOTSOFAR-1 challenge: Highlights and learn- ings,

    I. Abramovski, A. Vinnikov, S. Shaer, N. Kanda, X. Wang et al., “Summary of the NOTSOFAR-1 challenge: Highlights and learn- ings,” arXiv preprint arXiv:2501.17304, 2025

    2025 arXiv

  16. [24]

    Si- multaneous diarization and separation of meetings through the integration of statistical mixture models,

    T. Cord-Landwehr, C. Boeddeker, and R. Haeb-Umbach, “Si- multaneous diarization and separation of meetings through the integration of statistical mixture models,” arXiv preprint arXiv:2410.21455, 2024

    2024 arXiv

  17. [25]

    The generalized correlation method for estimation of time delay,

    C. Knapp and G. Carter, “The generalized correlation method for estimation of time delay,” IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 24, no. 4, pp. 320–327, 1976

    1976

  18. [26]

    Disambiguation of TDOA estimation for multiple sources in reverberant environments,

    J. Scheuing and B. Yang, “Disambiguation of TDOA estimation for multiple sources in reverberant environments,”IEEE Transac- tions on Audio, Speech, and Language Processing, vol. 16, no. 8, pp. 1479–1489, 2008

    2008

  19. [27]

    On the approximate W-disjoint or- thogonality of speech,

    S. Rickard and O. Yilmaz, “On the approximate W-disjoint or- thogonality of speech,” in Proc. IEEE ICASSP, vol. 1, 2002, pp. I–529–I–532

    2002

  20. [28]

    On optimal frequency- domain multichannel linear filtering for noise reduction,

    M. Souden, J. Benesty, and S. Affes, “On optimal frequency- domain multichannel linear filtering for noise reduction,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 18, no. 2, pp. 260–276, 2010

    2010

  21. [29]

    Multiple sound source counting and localization based on tf-wise spatial spectrum clus- tering,

    B. Yang, H. Liu, C. Pang, and X. Li, “Multiple sound source counting and localization based on tf-wise spatial spectrum clus- tering,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 27, no. 8, pp. 1241–1255, 2019

    2019

  22. [30]

    Correlation matrix distance, a meaningful measure for evaluation of non- stationary mimo channels,

    M. Herdin, N. Czink, H. Ozcelik, and E. Bonek, “Correlation matrix distance, a meaningful measure for evaluation of non- stationary mimo channels,” in IEEE 61st Vehicular Technology Conference, vol. 1, 2005, pp. 136–140 V ol. 1

    2005

  23. [31]

    Once more diarization: Improving meeting transcription systems through segment-level speaker reassignment,

    C. Boeddeker, T. Cord-Landwehr, and R. Haeb-Umbach, “Once more diarization: Improving meeting transcription systems through segment-level speaker reassignment,” in Proc. ISCA In- terspeech, 2024, pp. 1615–1619

    2024

  24. [32]

    Hi- erarchical density estimates for data clustering, visualization, and outlier detection,

    R. J. G. B. Campello, D. Moulavi, A. Zimek, and J. Sander, “Hi- erarchical density estimates for data clustering, visualization, and outlier detection,” ACM Trans. Knowl. Discov. Data, 2015

    2015

  25. [33]

    Continuous speech separation: Dataset and analysis,

    Z. Chen, T. Yoshioka, L. Lu, T. Zhou, Z. Menget al., “Continuous speech separation: Dataset and analysis,” in Proc. IEEE ICASSP, 2020, pp. 7284–7288

    2020

  26. [34]

    ESPnet2 pretrained model, Shinji Watan- abe/librispe ech asr train asr transformer e18 raw bpe sp valid .acc.best, fs=16k, lang=en,

    S. Watanabe, “ESPnet2 pretrained model, Shinji Watan- abe/librispe ech asr train asr transformer e18 raw bpe sp valid .acc.best, fs=16k, lang=en,” Jul. 2020. [Online]. Available: https://doi.org/10.5281/zenodo.3966501

    2020 doi

  27. [35]

    Front-end processing for the CHiME-5 dinner party scenario,

    C. Boeddecker, J. Heitkaemper, J. Schmalenstroeer, L. Drude, J. Heymann et al., “Front-end processing for the CHiME-5 dinner party scenario,” in Proc. 5th International Workshop on Speech Processing in Everyday Environments (CHiME), 2018, pp. 35–40

    2018

  28. [36]

    NIST 2021 speaker recognition evaluation plan,

    O. Sadjadi, C. Greenberg, E. Singer, L. Mason, and D. Reynolds, “NIST 2021 speaker recognition evaluation plan,” 2021-07-12 04:07:00 2021

    2021

  29. [37]

    MeetEval: A toolkit for computation of word error rates for meeting transcription systems,

    T. von Neumann, C. Boeddeker, M. Delcroix, and R. Haeb- Umbach, “MeetEval: A toolkit for computation of word error rates for meeting transcription systems,” in Proc. 7th Interna- tional Workshop on Speech Processing in Everyday Environments (CHiME), 2023, pp. 27–32

    2023

  30. [38]

    Multi-class spectral clus- tering with overlaps for speaker diarization,

    D. Raj, Z. Huang, and S. Khudanpur, “Multi-class spectral clus- tering with overlaps for speaker diarization,” inIEEE Spoken Lan- guage Technology Workshop (SLT), 2021, pp. 582–589

    2021

  31. [39]

    Geodesic interpolation of frame-wise speaker embeddings for the diarization of meeting scenarios,

    T. Cord-Landwehr, C. Boeddeker, C. Zoril ˘a, R. Doddipatla, and R. Haeb-Umbach, “Geodesic interpolation of frame-wise speaker embeddings for the diarization of meeting scenarios,” in Proc. IEEE ICASSP, 2024, pp. 11 886–11 890

    2024

  32. [40]

    GPU-accelerated guided source separation for meeting transcription,

    D. Raj, D. Povey, and S. Khudanpur, “GPU-accelerated guided source separation for meeting transcription,” in Proc. ISCA Inter- speech, 2023, pp. 3507–3511

    2023

  33. [41]

    V oxceleb: Large-scale speaker verification in the wild,

    A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “V oxceleb: Large-scale speaker verification in the wild,” Computer Science and Language, 2019

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.