Pith. sign in

REVIEW 4 major objections 7 minor 39 references

SDBench: A Comprehensive Benchmark Suite for Speaker Diarization

T0 review · 4 major / 7 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 A 13-dataset benchmark drives a 9.6x speedup for on-device speaker diarization with comparable error rates to the Pyannote baseline.

desk verdict SDBench is a genuinely useful benchmark suite, but the 9.6x speedup claim for SpeakerKit is not backed by the paper's own ablations. read the letter →

arxiv 2507.16136 v2 pith:JV3AZE4W submitted 2025-07-22 cs.SD cs.AIeess.AS

Eduardo Pacheco , Atila Orhon , Berkin Durmus , Blaise Munyampirwa , Andrey Leonov This is my paper
classification cs.SDcs.AIeess.AS
keywords speakerdiarizationbenchmarksuiteon-deviceinferenceembeddingablationstudyerrorratespeedPyannote
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes SDBench, an open-source benchmark suite for speaker diarization that integrates 13 diverse datasets under consistent evaluation conventions, including a fine-grained stage-wise error analysis. To demonstrate the suite's value, the authors use it to build SpeakerKit, a diarization system built on top of Pyannote v3.1, and report that ablation studies enabled by SDBench led to SpeakerKit running 9.6x faster than Pyannote while achieving comparable diarization error rates. The speedup comes from a per-chunk speaker embedding strategy that removes redundant forward passes and from widening the sliding-window stride with little measured accuracy loss. If the claim holds, near-real-time speaker diarization becomes practical on local devices without sacrificing the accuracy previously associated with heavy cloud compute.

What carries the argument

The central mechanism is the per-chunk speaker embedding strategy: instead of feeding each speaker's sliding windows to the embedding model separately, the model processes the whole audio chunk in one forward pass, and the speaker masks in the final temporal-statistics aggregation layer reconstruct the same per-speaker, per-window embeddings. The second mechanism is the sliding-window stride ablation, where the stride of the segmenter's window is increased from 1 to 2 to 4 seconds, trading temporal redundancy for speed while keeping DER nearly constant. SDBench itself provides the tooling for stage-wise evaluation through two idealized systems, Oracle Segmenter and Oracle Clusterer, which isolate segmentation error from clustering error.

What would settle it

Run both the per-window and per-chunk embedding strategies on the same audio chunks with the same model weights and compare the resulting embeddings element-wise, or by cosine distance; if any meaningful difference appears, the two strategies are not exactly equivalent and the claimed DER parity would need to be re-measured.

Watch

Extended reading notes

Core claim

The authors claim that the dominant computational cost in Pyannote v3.1 is redundancy in the speaker embedding stage, which runs the embedding model separately for each speaker over highly overlapping 10-second windows, producing a 21x overhead. They discover that by running the embedding model once on the entire audio chunk and recovering per-speaker, per-window embeddings through the final layer's speaker-mask pooling, the same sliding-window pattern is reproduced with no redundant computation, yielding a 1.2x embedding-stage speedup. Combined with their finding that increasing the sliding-window stride from 1 to 4 seconds reduces DER by at most 0.02 for up to 5 speakers and 0.05 for unlimited speakers, SpeakerKit achieves a 9.6x speedup over Pyannote v3.1 while keeping DER comparable. They also benchmark six systems, including Deepgram, AWS Transcribe, and the Pyannote AI API, and find that the Pyannote API has the lowest DER while SpeakerKit has the highest speed factor among locally run systems.

Load-bearing premise

SpeakerKit's error-rate parity with Pyannote depends on the per-chunk embedding pass producing exactly the same speaker embeddings as Pyannote's per-window pass, because only the inference pattern changes and not the model being simulated.

Editorial extensions

If this is right

  • If SpeakerKit's speedup holds, a 30-second audio clip can be diarized in roughly 3 seconds on a local device, making real-time on-device speaker diarization feasible for meeting scribes and voice assistants.
  • The stage-wise error breakdown lets practitioners target the stage that dominates error for their particular domain instead of treating the diarization system as an opaque pipeline.
  • The small DER impact of a 4-second stride suggests that systems deployed in controlled or structured domains can safely reduce sliding-window redundancy to gain large speedups.
  • Consistent, collar-free DER reporting across 13 datasets makes cross-system and cross-dataset comparisons more reproducible than the ad-hoc benchmarks common in the field.
  • Measuring server-side systems by total request completion time, including upload and polling, reveals end-to-end latency trade-offs that model-only benchmarks miss.

Reading between the lines

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

  • The per-chunk embedding strategy, because it relies only on final-layer speaker-mask pooling, could presumably be adopted inside Pyannote itself, potentially distributing the gain to all Pyannote users and not just SpeakerKit.
  • The 9.6x speedup is hardware-conditional, measured on an M2 Ultra with MPS and float16; on other backends the numeric equivalence of the per-chunk masks may need revalidation before the speedup is assumed portable.
  • The benchmark's finding that overlap-heavy datasets like AliMeetings remain challenging suggests that future benchmark extensions could add overlapped-speech handling as a first-class metric, a dimension SDBench currently only reports as a dataset statistic.
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 / 7 minor

Summary. The paper introduces SDBench, an open-source benchmark suite for speaker diarization that integrates 13 datasets and provides tooling for consistent DER and speed-factor evaluation. The authors use SDBench to compare six diarization systems and to run ablation studies on Pyannote v3.1, including stage-wise oracle evaluations, sliding-window stride variations, and a per-chunk speaker-embedding strategy. As a demonstration of the benchmark's usefulness, they present SpeakerKit, a modified Pyannote v3.1 system that uses per-chunk embeddings while maintaining stride 1, and claim it is 9.6x faster than Pyannote v3.1 with comparable DER.

Significance. If the claims are substantiated, SDBench would be a valuable community resource by addressing the lack of reproducible, multi-domain diarization evaluation. The per-chunk speaker-embedding idea is a practical and potentially generalizable optimization, and the stage-wise oracle decompositions provide useful insights into where error reductions can come from. The paper's strengths include open release of code and dataset-processing scripts, a diverse set of 13 datasets, and comparisons against external baselines without obvious circularity. However, the headline 9.6x speedup is not currently backed by the paper's own ablation data, and measurement uncertainty is not reported, so the practical significance is conditional on a transparent and reproducible measurement breakdown.

major comments (4)
  1. [Abstract, §2.4, Figure 1(a), Table 1] The central claim that SpeakerKit is 9.6x faster than Pyannote v3.1 is not derivable from the ablation evidence in the paper. Table 1 shows that the only adopted change (per-chunk embedding) yields speed factors of 446 versus 373, a 1.20x improvement, and Section 2.4 states that SpeakerKit maintains stride 1 for the speaker segmenter. The 15.3x to 38.3x speedups in Figure 3(a) are explicitly for stride 4, which SpeakerKit does not use. No measurement protocol, stage-level timing table, or supplementary breakdown is provided to explain how a 1.2x embedding-stage gain translates to a 9.6x end-to-end speedup. If the 9.6x figure relies on an unreported configuration change, the DER-comparability claim is also threatened, since Figure 3(a) shows DER penalties up to 0.06 at stride 4 on VoxConverse. Please provide a full and explicit timing breakdown of SpeakerKit versus Pyannote v3.1, or revise the claim to match the measurements actually reported.
  2. [§2.3, §4] The DER-parity argument rests on the assertion that the per-chunk speaker embedding strategy 'reproduce[s] the same sliding window pattern per speaker in the final layer by using the speaker masks.' The manuscript does not demonstrate that the per-chunk forward pass produces embeddings that are identical, or even within a controlled tolerance, to Pyannote's per-window strategy. Differences could arise from batch processing, floating-point numerics, or implementation subtleties in the final-layer mask aggregation. Because the 'comparable error rates' claim depends on this equivalence, please provide an embedding-level comparison (e.g., cosine similarity or downstream DER on a held-out set) or, failing that, per-dataset DER comparisons between per-window and per-chunk strategies with confidence intervals.
  3. [§2.2, Figure 1, Figure 3, Table 1] All DER and Speed Factor results are reported as point estimates with no error bars, confidence intervals, or description of the measurement protocol. The paper does not state the number of repeated runs, warmup procedures, hardware state, batch sizes, or how server-side network latency was averaged. Since the paper's conclusions are quantitative system comparisons (Figure 1b) and ablation trade-offs (Figure 3, Table 1), the absence of variance information makes it impossible to judge which differences are meaningful. Please report multiple runs for local systems and multiple request samples for server-side APIs, with confidence intervals, or explicitly justify why the measurements are deterministic.
  4. [§3, §2.4] The statement that SpeakerKit achieves 'comparable DER' to Pyannote v3.1 is not supported by numeric values or a statistical equivalence test. Figure 1(a) shows a visual comparison, but the paper does not report the actual DER figures for SpeakerKit and Pyannote v3.1, nor confidence intervals on the DER difference. Since the abstract's claim depends on this parity, please provide a table of per-dataset DER values for the two systems and a quantitative statement of what 'comparable' means, such as a maximum acceptable DER degradation or a confidence interval on the difference.
minor comments (7)
  1. [§3, first paragraph] The text contains a typo: 'SDBenh' should be 'SDBench'.
  2. [§2.4, system list] The system list has a typo: 'A WS Transcribe' should be 'AWS Transcribe'.
  3. [Throughout] There are several spacing artifacts in names such as 'V oxConverse' and 'AliMetings'; these should be corrected to 'VoxConverse' and 'AliMeetings' for consistent presentation.
  4. [Figure 3(a), §2.2] The axis label 'Speed-up in pyannote units' is unclear; please define this unit explicitly, for example as the speed factor relative to Pyannote v3.1 on the same dataset.
  5. [Reference [27]] Reference [27] is listed as 'Whiperkit' but the URL points to WhisperKit; please correct the spelling and ensure the reference text matches the actual project name.
  6. [§2.1] The Hugging Face link is said to 'be available in the camera-ready version'; for a benchmark paper, the dataset-processing scripts and the benchmark harness should be made available at submission time in a clearly versioned form, rather than deferred to a future revision.
  7. [Figure 1(b)] Because the plot is on a log scale, consider adding a table with per-system aggregate DER and Speed Factor values so readers can extract exact numbers without reading the figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are external, measured benchmark results; SDBench's use as an ablation driver is a self-improvement loop, not a derivation from its own inputs.

full rationale

The paper's load-bearing claims—comparative DER and Speed Factor across six diarization systems, the 9.6x SpeakerKit speedup, and the 1.2x per-chunk embedding gain—are empirical measurements against external baselines (Pyannote v3.1, AWS Transcribe, Deepgram, Picovoice, Pyannote AI API) on externally defined datasets and metrics, not consequences of assumptions that include those results. The design claim that the per-chunk embedding 'reproduce[s] the same sliding window pattern per speaker in the final layer by using the speaker masks' (Section 2.3) is an implementation-equivalence assertion tested by the measured DER difference of 0.002, not an equation that forces the outcome. The only self-citation is Argmax's WhisperKit [27] used to predict dataset language for per-language breakdowns; that labeling step does not determine the central DER or speed comparisons. The skeptic's observation that the 9.6x figure is not directly derivable from Table 1's 1.2x ablation is an evidence-traceability concern, not circularity: no parameter is fitted to the benchmark result and no prior result is imported from the authors to forbid alternatives. SDBench's efficacy is demonstrated by the system it guided, which is a credibility claim rather than a circular derivation. Under the stated review rules, the derivation chain is self-contained against external benchmarks, so score 0 is appropriate.

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

This is a benchmark and systems paper, so it introduces no new physical or mathematical entities and fits no free parameters. Its central claims rest on measurement assumptions about datasets, metrics, and latency definitions.

assumptions (4)
  • domain assumption Ground truth speaker labels in all 13 datasets are accurate and aligned with the provided RTTM files.
    The entire DER computation depends on the correctness of ground truth annotations, which is asserted without verification in Section 2.1.
  • domain assumption DER without forgiveness collars is the appropriate metric for apples-to-apples comparison.
    The paper deliberately omits collars (Section 2.2), assuming this gives a fair cross-system comparison, though many prior benchmarks use collars.
  • domain assumption Whisper large-v3 language predictions are accurate enough for per-language evaluation.
    Whisper is used to label each file's language in Section 2.1, and incorrect labels would distort the language breakdown in Figure 5.
  • domain assumption Server-side API latency should include upload, polling, and download time.
    The Speed Factor definition in Section 2.2 includes these overheads, which is a fair design choice but not the only way to measure server-side speed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SDBench: A Comprehensive Benchmark Suite for Speaker Diarization." pith.science (2026). https://pith.science/paper/JV3AZE4W

@misc{pith2026250716136,
  author       = {Pith},
  title        = {Pith review of: SDBench: A Comprehensive Benchmark Suite for Speaker Diarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JV3AZE4W}},
  note         = {Machine review of arXiv:2507.16136}
}
read the original abstract

Even state-of-the-art speaker diarization systems exhibit high variance in error rates across different datasets, representing numerous use cases and domains. Furthermore, comparing across systems requires careful application of best practices such as dataset splits and metric definitions to allow for apples-to-apples comparison. We propose SDBench (Speaker Diarization Benchmark), an open-source benchmark suite that integrates 13 diverse datasets with built-in tooling for consistent and fine-grained analysis of speaker diarization performance for various on-device and server-side systems. SDBench enables reproducible evaluation and easy integration of new systems over time. To demonstrate the efficacy of SDBench, we built SpeakerKit, an inference efficiency-focused system built on top of Pyannote v3. SDBench enabled rapid execution of ablation studies that led to SpeakerKit being 9.6x faster than Pyannote v3 while achieving comparable error rates. We benchmark 6 state-of-the-art systems including Deepgram, AWS Transcribe, and Pyannote AI API, revealing important trade-offs between accuracy and speed.

Figures

Figures reproduced from arXiv: 2507.16136 by the authors.

Figure 1
Figure 1. DER vs Speed Factor: (a) Pyannote v3.1 and Speak￾erKit; SpeakerKit demonstrates a 9.6x speedup while achieving comparable DER, (b) 6 speaker diarization systems across 13 datasets. Circle markers represent per-dataset results whereas star markers represent cross-dataset aggregation. *These authors contributed equally. 1The code is available at https://github.com/ argmaxinc/SDBench 1. Introduction and Related Work Sp… view at source ↗
Figure 2
Figure 2. presents key statistics of the test splits of each dataset: (a) total audio length, (b) overlap ratio, (c) speaker congestion, and (d) median speaker count. ICSI contains the longest total audio duration at 70 hours, while CALLHOME [15] and AMI [20] variants contain less than 10 hours each. AliMeetings [22] shows the highest overlap ratio (0.19) among all datasets, indicating frequent speaker overlaps, while Earning… view at source ↗
Figure 3
Figure 3. Sliding window strategy. (a) Relation between DER and Speed-up in pyannote units for strides 1, 2 and 4. (b) Shows the DER comparison between stride 1 and 4 across speaker counts. DER is impacted by less than 0.02 for up to 5 speak￾ers and by less than 0.05 for unlimited number of speakers. defined in Pyannote [4]. DER is calculated as follows: DER = False alarm + Missed detection + Speaker confusion Time Duration (… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Stage-wise evaluations. Oracle Segmenter and Or￾acle Clusterer systems are compared to the standard Pyannote v3.1 system (Base) without any idealized components across 5 datasets. DER is broken across Missed Detection, False Alarm, and Confusion. window inference which…
Figure 5
Figure 5. Figure 5: DER breakdown across multiple systems, categorized by language. 6. References [1] N. Ryant, P. Singh, V. Krishnamohan, R. Varma, K. Church, C. Cieri, J. Du, S. Ganapathy, and M. Liberman, “The third dihard diarization challenge,” 2021. [Online]. Available: https://arxi…
Figure 6
Figure 6. Figure 6: DER breakdown across multiple systems, categorized by number of speakers. [5] Amazon Web Services, “Amazon transcribe,” 2023, automatic speech recognition service. [Online]. Available: https://aws. amazon.com/transcribe/ [6] Pyannote, “pyannoteai api,” speaker Diarizat…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 32 canonical work pages

  1. [1]

    Introduction and Related Work Speaker diarization, the task of identifying ”who spoke when” in audio data, is critical for applications such as meeting tran- scriptions, scribes, and voice assistants. Top-ranking systems on several recent challenges ([1], [2], [3]) have adopted the following multi-stage architecture: (i) speaker segmentation divides local...

    arXiv 2025

  2. [2]

    Experiments 2.1. Datasets SDBench integrates 13 datasets spanning public and propri- etary sources that represent multiple languages, audio domains, and speaker distributions: CALLHOME [15], DIHARD-III [16], ICSI [17], Earnings-21 [18], MSDWild [19], AMI-IHM [20], AMI-SDM [20], V oxConverse [21], AliMetings [22], AISHELL-4 [23], American-Life-Podcast [24]...

    work page

  3. [3]

    Figure 4 breaks down DER into 3 components: (i) Missed Detection, (ii) False Alarm, and (iii) Confusion

    Results Stage-wise evaluations. Figure 4 breaks down DER into 3 components: (i) Missed Detection, (ii) False Alarm, and (iii) Confusion. (i) and (ii) are attributed to the speaker segmentation stage while (iii) is attributed to the clustering stage. These results demonstrate that different error types are dominant across different datasets. However, the f...

    work page

  4. [4]

    To demonstrate its effectiveness, we developed SpeakerKit, an inference efficiency-optimized system built on Pyannote v3 that maintains low DER while achieving a 9.6x speedup

    Discussion & Conclusions We introduced SDBench, an open-source speaker diarization benchmark toolkit enabling fine-grained error analysis across various domains through stage-wise evaluations and 13 di- verse datasets. To demonstrate its effectiveness, we developed SpeakerKit, an inference efficiency-optimized system built on Pyannote v3 that maintains lo...

    work page

  5. [5]

    Appendix Figure 5: DER breakdown across multiple systems, categorized by language

    work page

  6. [6]

    pyannoteai api,

    Pyannote, “pyannoteai api,” speaker Diarization API. [Online]. Available: https://pyannote.ai/

    work page

  7. [7]

    gladia api,

    Gladia, “gladia api,” speaker Diarization API. [Online]. Available: https://app.gladia.io/

    work page

  8. [8]

    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,” 2021. [Online]. Available: https://arxiv.org/abs/2012.01477

    arXiv 2021

Show all 39 references
  1. [9]

    The voxceleb speaker recognition challenge: A retrospective,

    J. Huh, J. S. Chung, A. Nagrani, A. Brown, J.-w. Jung, D. Garcia-Romero, and A. Zisserman, “The voxceleb speaker recognition challenge: A retrospective,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 32, p. 3850–3866, 2024. [Online]. Available: http://...

    2024

  2. [10]

    Notsofar-1 challenge: New datasets, baseline, and tasks for distant meeting transcription,

    A. Vinnikov, A. Ivry, A. Hurvitz, I. Abramovski, S. Koubi, I. Gurvich, S. Pe‘er, X. Xiao, B. M. Elizalde, N. Kanda, X. Wang, S. Shaer, S. Yagev, Y . Asher, S. Sivasankaran, Y . Gong, M. Tang, H. Wang, and E. Krupka, “Notsofar-1 challenge: New datasets, baseline, and tasks for ...

    2024 arXiv

  3. [11]

    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 Interspeech 2023 , 2023, pp. 3222–3226. Figure 6: DER breakdown across multiple systems, categorized by number of speakers

    2023

  4. [12]

    Amazon transcribe,

    Amazon Web Services, “Amazon transcribe,” 2023, automatic speech recognition service. [Online]. Available: https://aws. amazon.com/transcribe/

    2023

  5. [13]

    and VBx [14] as they adopt SDBench

  6. [14]

    Vbx: Speaker verification using bayesian inference,

    F. Landini et al. , “Vbx: Speaker verification using bayesian inference,” in ICASSP 2021 , 2021, pp. 6149–6153. [Online]. Available: https://github.com/BUTSpeechFIT/VBx

    2021

  7. [15]

    and AMI [20] variants contain less than 10 hours each. AliMeetings [22] shows the highest overlap ratio (0.19) among all datasets, indicating frequent speaker overlaps, while Earnings-21 [18] and American-Life-Podcast [24] show minimal overlap, reflecting their more structured...

  8. [16]

    Deepgram,

    Deepgram, “Deepgram,” 2023, speech tasks API platform. [Online]. Available: https://www.deepgram.com/

    2023

  9. [17]

    Picovoice sdk,

    P. Inc., “Picovoice sdk,” 2023. [Online]. Available: https: //github.com/Picovoice/falcon/tree/1.0.1

    2023

  10. [18]

    Picovoice diarization benchmarks

    Picovoice, “Picovoice diarization benchmarks.” [Online]. Available: https://github.com/Picovoice/ speaker-diarization-benchmark/

  11. [19]

    pyannote github issue

    pyannote, “pyannote github issue.” [Online]. Available: https: //github.com/pyannote/pyannote-audio

  12. [20]

    Pyannote github issue

    Pyannote, “Pyannote github issue.” [Online]. Available: https: //github.com/pyannote/pyannote-audio/issues/1634

  13. [21]

    Nemo open source speaker diarization system

    T. Park, N. R. Koluguri, F. Jia, J. Balam, and B. Ginsburg, “Nemo open source speaker diarization system.” in INTERSPEECH, 2022, pp. 853–854

    2022

  14. [22]

    Callhome american english speech,

    Canavan, Alexandra, Graff, David, and Zipperlen, George, “Callhome american english speech,” 1997. [Online]. Available: https://catalog.ldc.upenn.edu/LDC97S42

    1997

  15. [23]

    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,” inInterspeech 2021, 2021, pp. 3570–3574

    2021

  16. [24]

    The icsi meeting corpus,

    A. Janin, D. Baron, J. Edwards, D. Ellis, D. Gelbart, N. Morgan, B. Peskin, T. Pfau, E. Shriberg, A. Stolcke, and C. Wooters, “The icsi meeting corpus,” in 2003 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2003. Proceedings. (ICASSP ’03). , ser. I...

    2003 arXiv

  17. [25]

    Earnings-21: A practical benchmark for asr in the wild,

    M. Del Rio, N. Delworth, R. Westerman, M. Huang, N. Bhandari, J. Palakapilly, Q. McNamara, J. Dong, P. ˙Zelasko, and M. Jett ´e, “Earnings-21: A practical benchmark for asr in the wild,” in Interspeech 2021 . ISCA, Aug. 2021, p. 3465–3469. [Online]. Available: http://dx.doi.or...

    2021 doi

  18. [26]

    Msdwild: Multi-modal speaker diarization dataset in the wild,

    T. Liu, S. Fan, X. Xiang, H. Song, S. Lin, J. Sun, T. Han, S. Chen, B. Yao, S. Liu, Y . Wu, Y . Qian, and K. Yu, “Msdwild: Multi-modal speaker diarization dataset in the wild,” in Interspeech 2022. ISCA, Sep. 2022, p. 1476–1480. [Online]. Available: http://dx.doi.org/10.21437/...

    2022 doi

  19. [27]

    Announcing the ami meeting corpus,

    J. Carletta, “Announcing the ami meeting corpus,” The ELRA Newsletter, vol. 11, no. 1, pp. 3–5, 2006

    2006

  20. [28]

    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,” in Interspeech 2020 . ISCA, Oct. 2020. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2020-2337

    2020 doi

  21. [29]

    M2met: The icassp 2022 multi-channel multi-party meeting transcription challenge,

    F. Yu, S. Zhang, Y . Fu, L. Xie, S. Zheng, Z. Du, W. Huang, P. Guo, Z. Yan, B. Ma, X. Xu, and H. Bu, “M2met: The icassp 2022 multi-channel multi-party meeting transcription challenge,”

    2022

  22. [30]

    Wespeaker: A research and production oriented speaker embedding learning toolkit,

    H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y . Deng, and Y . Qian, “Wespeaker: A research and production oriented speaker embedding learning toolkit,” in IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  23. [31]

    Aishell-4: An open source dataset for speech enhancement, separation, recognition and speaker diarization in conference scenario,

    Y . Fu, L. Cheng, S. Lv, Y . Jv, Y . Kong, Z. Chen, Y . Hu, L. Xie, J. Wu, H. Bu, X. Xu, J. Du, and J. Chen, “Aishell-4: An open source dataset for speech enhancement, separation, recognition and speaker diarization in conference scenario,” 2021. [Online]. Available: https://a...

    2021 arXiv

  24. [32]

    Speech recognition and multi-speaker diarization of long conversations,

    H. H. Mao, S. Li, J. McAuley, and G. Cottrell, “Speech recognition and multi-speaker diarization of long conversations,”

  25. [34]

    Ava-avd: Audio-visual speaker diarization in the wild,

    E. Z. Xu, Z. Song, S. Tsutsui, C. Feng, M. Ye, and M. Z. Shou, “Ava-avd: Audio-visual speaker diarization in the wild,” in Proceedings of the 30th ACM International Conference on Multimedia . ACM, Oct. 2022, p. 3838–3847. [Online]. Available: http://dx.doi.org/10.1145/3503161.3548027

    2022

  26. [35]

    Ego4d: Around the world in 3,000 hours of egocentric video,

    K. Grauman, A. Westbury, L. Torresani, K. Kitani, J. Malik, T. Afouras et al. , “Ego4d: Around the world in 3,000 hours of egocentric video,” arXiv preprint arXiv:2204.01640, 2022

    2022 arXiv

  27. [36]

    Whiperkit

    Argmax, “Whiperkit.” [Online]. Available: https://github.com/ argmaxinc/WhisperKit

  28. [37]

    Wavlm: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,” IEEE Journal of Selecte...

    2022

  29. [38]

    Huggingface

    HuggingFace, “Huggingface.” [Online]. Available: https:// huggingface.co/

  30. [2020]

    Available: https://arxiv.org/abs/2005.08072

    [Online]. Available: https://arxiv.org/abs/2005.08072

    2005 arXiv

  31. [2022]

    Available: https://arxiv.org/abs/2110.07393

    [Online]. Available: https://arxiv.org/abs/2110.07393

    arXiv

Pith tools

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