Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Error Analysis in a Modular Meeting Transcription System

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Missing speech segments, not cross-channel leakage, are the main cause of the remaining word-error gap in a strong modular meeting transcription system.

desk verdict Missing segments, not leakage, dominate the gap to oracle segmentation; the analysis is useful but the oracle heuristics are under-specified for the exact numbers to be fully trusted. read the letter →

arxiv 2509.10143 v1 pith:D2XE3M4H submitted 2025-09-12 eess.AS cs.CLcs.LGcs.SD

Peter Vieting , Simon Berger , Thilo von Neumann , Christoph Boeddeker , Ralf Schlüter , Reinhold Haeb-Umbach This is my paper · ORCID
classification eess.AScs.CLcs.LGcs.SD
keywords meetingtranscriptionspeechseparationleakageanalysissegmentationerrorsvoiceactivitydetectiondiarizationLibriCSScoincidencerate
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 argues that in a modular meeting transcription pipeline—continuous speech separation followed by segmentation, ASR, and diarization—the largest remaining source of error is not leakage of speech between separated channels but missing speech segments. The authors extend a frame-wise alignment-based leakage analysis to capture primary-to-cross leakage in regions where only the primary speaker is active, and show that although such leakage is measurable and substantial, the voice activity detector mostly ignores it, so it does little damage. Using oracle-informed heuristics that eliminate one segmentation error type at a time, they find that 'missing' segments—speech that the segmentation drops entirely—account for more than half of the word-error-rate gap to oracle segmentation. Replacing the simple energy-based VAD with a transcription-supported diarization closes about a third of that gap and yields state-of-the-art single-microphone results on LibriCSS among systems that train the recognizer on LibriSpeech only. The paper's value is diagnostic: it redirects attention from separation leakage toward segmentation recall.

What carries the argument

The central objects are frame-wise coincidence rates (CRs) computed from word-level alignments and lattices, extended to measure primary-to-cross-channel leakage, and oracle-informed heuristics that surgically eliminate each segmentation error type (leakage, missing, merges, boundaries) to isolate its contribution to word error rate.

What would settle it

A concrete check: run the same ASR pipeline with a VAD that is artificially given oracle knowledge of all speech presence (or with a lower energy threshold that recovers missing segments). If the resulting ORC WER does not improve by roughly the 0.4–0.7 percentage points attributed to 'missing,' the oracle-informed heuristic overstates that error type's contribution.

Watch

Extended reading notes

Core claim

The central claim is that segmentation recall—not cross-channel leakage—limits a strong modular meeting transcription system. Extending an earlier leakage-analysis framework with coincidence rates computed between forced alignments on the primary channel and hypotheses on the cross channel, the paper measures substantial primary-to-cross leakage in frames where only the primary speaker is active; yet the VAD does not create segments on those leaked regions, so the effect on final WER is negligible. The decisive result is the segmentation error breakdown: a heuristic that removes 'missing' segments improves ORC WER by 0.4–0.7 percentage points absolute, and this error type accounts for more t

Load-bearing premise

The segmentation-error attribution assumes that each error type can be eliminated one at a time without changing the segmentation in other ways, so the measured WER difference cleanly belongs to that error type.

Editorial extensions

If this is right

  • Diarization using ASR timestamps reduces the ORC WER gap to oracle segmentation by about one third compared with an energy-based VAD.
  • Missing segments are the dominant segmentation error; eliminating them improves ORC WER by 0.4–0.7 percentage points absolute and accounts for more than half of the gap.
  • Primary-to-cross-channel leakage, though measurable in single-speaker regions, is largely ignored by VAD and does not materially change the final recognition result.
  • Fixing leakage, merges, and boundaries together leaves the WER nearly unchanged, so improving VAD recall is the most direct path to better meeting transcription.
  • The system reaches the best single-microphone LibriCSS performance among systems that train ASR on LibriSpeech data only.

Reading between the lines

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

  • Because the attribution heuristics assume error types can be removed independently, the 'missing segments' share may be inflated if removing a segment also shifts boundaries or alters language-model context; a controlled experiment that adds back random oracle segments would test this.
  • An immediate testable extension: replace the energy VAD with a high-recall VAD (e.g., tuned with lower thresholds) and measure whether the WER improvement matches the predicted 'missing' contribution; if it does not, the heuristic is confounded.
  • The same frame-wise coincidence framework could be applied to end-to-end systems or streaming setups, where leakage across context windows may manifest differently.
  • If missing segments dominate, gains from stronger separation models may be limited unless paired with segmentation improvements—relevant to future system design.
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 paper analyzes a modular single-microphone meeting transcription pipeline (continuous speech separation, energy-based VAD or ASR-supported diarization, hybrid ASR, optional second ASR pass) on LibriCSS. It extends a previously proposed leakage-analysis framework to measure primary-to-cross-channel leakage using frame-wise coincidence rates. The central empirical claims are: (i) there is significant leakage into the cross channel in regions where only the primary speaker is active; (ii) this leakage does not substantially affect final ORC WER; (iii) advanced diarization closes about a third of the gap to oracle segmentation compared with an energy-based VAD; and (iv) among the remaining segmentation error types, missing speech segments account for more than half of the gap to the oracle, whereas leakage, merges, and boundary deviations have small effects. The system is reported to achieve state-of-the-art single-microphone LibriCSS cpWER among systems that train the ASR module on LibriSpeech data only.

Significance. If the segmentation-error decomposition is valid, the paper gives a practically useful and falsifiable ranking of where to invest future effort (improved VAD to recover missing segments, rather than leakage suppression or boundary refinement). The extension of the leakage framework to primary-to-cross-channel leakage is a sensible and clearly motivated methodological contribution. The use of a public benchmark (LibriCSS) and the comparison of multiple segmentations are strengths. However, the central attribution of the oracle gap to missing segments rests on oracle-informed heuristics that are not specified in the manuscript and on small WER differences without uncertainty quantification. The state-of-the-art claim is credible within its stated scope but depends on a careful reading of Table 1. Overall, the paper is potentially valuable, but the main quantitative conclusion is not yet reproducible from the text.

major comments (3)
  1. [Section 3.2 and Table 4] The oracle-informed heuristics are the load-bearing tool of the paper, but they are described only by the sentence "For each error type, we define a heuristic that exploits oracle information to eliminate these errors from the segmentation." No algorithmic definitions are given. The central claim that missing segments "account for more than half of the performance gap" (Section 5.2) cannot be checked unless the authors specify exactly how each error type is removed and verify that removing one error type does not change others. The reported numbers are not fully additive: for Whisper/4gram, the individual reductions are 0.2+0.4+0.0+0.0=0.6 but "All" reduces by 0.8; for Ours/Trafo, individual reductions sum to 0.6 while "All" reduces by 0.7. This suggests interactions between the interventions, which must be analyzed or at least discussed. Please provide pseudocode or a precise descriptio
  2. [Section 5.2, Table 4] The WER differences used to rank error types are small (0.1 to 0.6% absolute) and no error bars, confidence intervals, or significance tests are reported. LibriCSS is a finite test set; the difference between "Missing" and "Merges" may be within run-to-run or test-set sampling noise. The claim that leakage, merges, and boundaries "do not have a significant effect" (Section 5.2) needs statistical support. Please report the number of sessions/utterances/words, and include bootstrap confidence intervals or a paired significance test for the ORC WER differences in Table 4.
  3. [Section 5.1, Table 3] The primary-to-cross-channel leakage measurement compares the primary channel forced alignment to 1-best or lattice hypotheses on the cross channel. In single-speaker regions the cross channel is expected to contain silence or leaked primary speech, but the ASR hypotheses on the cross channel could also contain hallucinated words or residual non-target speech. The paper reports high "words only" coincidence rates for one active speaker (8.2% for 1-best), but the evaluation protocol for these regions is not fully specified: how are "active speaker" regions defined, and are cross-channel segments passed through the same VAD/ASR pipeline? Please clarify the exact measurement procedure and discuss whether the measured coincidence could be inflated by ASR artifacts rather than true leakage.
minor comments (5)
  1. [Section 5, oracle segmentation description] "selecting the separated channel with minimum signal-to-distortion ratio (SDR) to the clean audio" should presumably read "maximum SDR", since higher SDR indicates better separation. If the minimum is actually intended, please explain why.
  2. [Table 1] The table is very hard to read: the row/column alignment is broken, and the rows containing "Oracle", "Ours", and the cpWER values are visually ambiguous. Please reformat so that each system, its training data, and the resulting cpWER are clearly aligned.
  3. [Figure 1] The example transcriptions in Figure 1 contain illegible tokens like "aa" and "nn". These appear to be formatting artifacts; please replace them with readable text.
  4. [Section 3.1] The sentence "The GCR computation for this example is shown by Figure 2" is confusing because Figure 2 uses different example words than Figure 3. Please align the examples or clarify the cross-reference.
  5. [Section 5.1] The statement that the CRs for primary hypotheses with cross-channel ground truth "are higher than the natural coincidence, but this is mainly caused by silence" should explicitly point to the "Words only" column in Table 3, since the reader must otherwise infer which numbers support the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper's claims are empirical ablations on an external benchmark.

full rationale

The paper's load-bearing conclusions are based on direct measurements on the LibriCSS test set and on oracle-informed ablations, not on fitted parameters or self-referential definitions. The leakage analysis computes coincidence rates between channel alignments, and the claim that leakage does not affect final performance much is tested in Section 5.2 by removing leaked segments and measuring WER changes; this is an interventionist ablation, not a prediction imposed by the metric. Similarly, the attribution of the performance gap to missing speech segments is an empirical comparison: eliminating each error type and measuring the ORC WER improvement does not force the result by construction, since the relative contribution of each error type depends on the actual ASR behavior on the removed/added audio. The paper reuses the authors' own system from [14] and diarization from [17]; these self-citations describe the experimental platform and prior published work that is externally established on LibriCSS, so they are not an unverified load-bearing premise. No uniqueness theorem or ansatz is smuggled in from same-author work. One genuine weakness is that the oracle-informed heuristics in Section 3.2 are not specified, making the Table 4 attribution hard to reproduce and potentially confounded; however, this is a missing-support/reproducibility issue, not a circular reduction, and therefore does not raise the circularity score.

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

The central empirical claims are benchmark measurements against oracle annotations, so they use no fitted target-derived parameters. The free parameters above are system hyperparameters that shape what the analysis measures, but they are not fit to the claimed conclusions.

free parameters (4)
  • VAD energy-ratio threshold = not reported
    Used to decide speech activity on the two separated channels (Section 2.2); tuned on dev Session0 individually for each diarization variant, directly shaping the segmentation error mix in Section 5.2.
  • Diarization speaker-change similarity threshold and context lengths (3s/4s) = not reported
    From the transcription-supported diarization of [17] (Section 2.4/4.2); determines when and where segments are split, influencing the missing/merges/boundaries error counts in Table 4.
  • CSS window size and shift = 4s / 3s
    Fixed in Section 4.2 to satisfy the at-most-two-speakers assumption; defines which leakage events are measurable at all.
  • Segment merging pause threshold = 3s
    Applied when rerunning ASR after diarization (Section 4.2); affects context length and thus the measured WER, including the oracle-gap numbers.
assumptions (4)
  • domain assumption At most two speakers are active within any 4s separation window
    CSS premise in Section 2.1; leakage analysis and the entire pipeline rely on it.
  • domain assumption Forced alignments from clean LibriSpeech signals, synchronized to LibriCSS, provide accurate word-level ground truth
    Section 3.1 replaces alignments on separated audio with clean alignments; all leakage coincidence rates depend on this synchronization.
  • domain assumption The oracle segmentation obtained by selecting the channel with minimum SDR to clean audio is the correct reference
    Used for oracle rows in Tables 1-2 and as the baseline for the error-gap decomposition; if oracle is imperfect, the gap attribution shifts.
  • domain assumption Word-level coincidence rates are a valid proxy for physical leakage
    Section 3.1 defines CR/GCR; the claim of 'significant leakage' is directly read from these proxy rates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Error Analysis in a Modular Meeting Transcription System." pith.science (2026). https://pith.science/paper/D2XE3M4H

@misc{pith2026250910143,
  author       = {Pith},
  title        = {Pith review of: Error Analysis in a Modular Meeting Transcription System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D2XE3M4H}},
  note         = {Machine review of arXiv:2509.10143}
}
read the original abstract

Meeting transcription is a field of high relevance and remarkable progress in recent years. Still, challenges remain that limit its performance. In this work, we extend a previously proposed framework for analyzing leakage in speech separation with proper sensitivity to temporal locality. We show that there is significant leakage to the cross channel in areas where only the primary speaker is active. At the same time, the results demonstrate that this does not affect the final performance much as these leaked parts are largely ignored by the voice activity detection (VAD). Furthermore, different segmentations are compared showing that advanced diarization approaches are able to reduce the gap to oracle segmentation by a third compared to a simple energy-based VAD. We additionally reveal what factors contribute to the remaining difference. The results represent state-of-the-art performance on LibriCSS among systems that train the recognition module on LibriSpeech data only.

Figures

Figures reproduced from arXiv: 2509.10143 by the authors.

Figure 1
Figure 1. Meeting transcription pipeline. A continuous speech separation (CSS) system separates overlapping speech of multiple [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Visualization of GCR computation measured between [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the leakage analysis from the cross chan [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 3 linked inside Pith

  1. [14]

    Combining TF-GridNet and mixture encoder for continuous speech separation for meeting transcription,

    P. Vieting, S. Berger, T. v. Neumann, C. Boeddeker, R. Schlüter, and R. Haeb-Umbach, “Combining TF-GridNet and mixture encoder for continuous speech separation for meeting transcription,” inProc. SLT, Macao, China, Dec. 2024, pp. 160–167

    2024

  2. [1]

    Continuous speech separation with conformer,

    S. Chen, Y . Wu, Z. Chen, J. Wu, J. Li, T. Yoshioka, C. Wang, S. Liu, and M. Zhou, “Continuous speech separation with conformer,” inProc. ICASSP, 2021, pp. 5749–5753

    2021

  3. [2]

    Mixture encoder for joint speech separation and recognition,

    S. Berger, P. Vieting, C. Boeddeker, R. Schlüter, and R. Haeb- Umbach, “Mixture encoder for joint speech separation and recognition,” inProc. Interspeech, Dublin, Ireland, Aug. 2023, pp. 3527–3531

    2023

  4. [3]

    Multi-turn rnn-t for streaming recognition of multi-party speech,

    I. Sklyar, A. Piunova, X. Zheng, and Y . Liu, “Multi-turn rnn-t for streaming recognition of multi-party speech,” in Proc. ICASSP, Singapore, May 2022, pp. 8402–8406

    2022

  5. [4]

    Streaming multi-talker ASR with token-level serialized output train- ing,

    N. Kanda, J. Wu, Y . Wu, X. Xiao, Z. Meng, X. Wang, Y . Gaur, Z. Chen, J. Li, and T. Yoshioka, “Streaming multi-talker ASR with token-level serialized output train- ing,” Preprint arXiv:2202.00842, 2022

    arXiv 2022

  6. [5]

    Continuous speech separation: Dataset and analysis,

    Z. Chen, T. Yoshioka, L. Lu, T. Zhou, Z. Meng, Y . Luo, J. Wu, X. Xiao, and J. Li, “Continuous speech separation: Dataset and analysis,” inProc. ICASSP, Barcelona, Spain, May 2020, pp. 7284–7288

    2020

  7. [6]

    TF-GridNet: Integrating full- and sub-band modeling for speech separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee, B.-Y . Kim, and S. Watanabe, “TF-GridNet: Integrating full- and sub-band modeling for speech separation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 3221–3236, 2023

    2023

  8. [7]

    TS-SEP: Joint diarization and separation conditioned on estimated speaker embeddings,

    C. Boeddeker, A. S. Subramanian, G. Wichern, R. Haeb- Umbach, and J. L. Roux, “TS-SEP: Joint diarization and separation conditioned on estimated speaker embeddings,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 1185–1197, 2024

    2024

Show all 28 references
  1. [8]

    M2Met: The ICASSP 2022 multi-channel multi-party meeting transcrip- tion 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 transcrip- tion challenge,” inProc. ICASSP, Singapore, May 2022, pp. 6167–6171

    2022

  2. [9]

    How bad are artifacts?: Analyzing the impact of speech enhancement errors on ASR,

    K. Iwamoto, T. Ochiai, M. Delcroix, R. Ikeshita, H. Sato, S. Araki, and S. Katagiri, “How bad are artifacts?: Analyzing the impact of speech enhancement errors on ASR,” inProc. Interspeech, Incheon, Korea, Sep. 2022, pp. 5418–5422

    2022

  3. [10]

    Impact of residual noise and artifacts in speech enhancement errors on intelligibility of human and machine,

    S. Araki, A. Yamamoto, T. Ochiai, K. Arai, A. Ogawa, T. Nakatani, and T. Irino, “Impact of residual noise and artifacts in speech enhancement errors on intelligibility of human and machine,” inProc. Interspeech, Dublin, Ireland, Aug. 2023, pp. 2503–2507

    2023

  4. [11]

    Monaural source separation: From anechoic to reverberant environments,

    T. Cord-Landwehr, C. Boeddeker, T. V on Neumann, C. Zo- ril˘a, R. Doddipatla, and R. Haeb-Umbach, “Monaural source separation: From anechoic to reverberant environments,” in Proc. IWAENC, Bamberg, Germany, 2022

    2022

  5. [12]

    Explicit word error minimization using word hypothesis posterior probabilities,

    F. Wessel, R. Schlüter, and H. Ney, “Explicit word error minimization using word hypothesis posterior probabilities,” inProc. ICASSP, Salt Lake City, UT, USA, May 2001, pp. 33–36

    2001

  6. [13]

    Con- fidence measures for large vocabulary continuous speech recognition,

    F. Wessel, R. Schlüter, K. Macherey, and H. Ney, “Con- fidence measures for large vocabulary continuous speech recognition,”IEEE Transactions on Speech and Audio Pro- cessing, vol. 9, no. 3, pp. 288–298, Mar. 2001

    2001

  7. [15]

    Recognizing overlapped speech in meetings: A multichan- nel separation approach using neural networks,

    T. Yoshioka, H. Erdogan, Z. Chen, X. Xiao, and F. Alleva, “Recognizing overlapped speech in meetings: A multichan- nel separation approach using neural networks,” inProc. Interspeech, Hyderabad, India, Sep. 2018, pp. 3038–3042

    2018

  8. [16]

    TF-GridNet: Making time-frequency domain models great again for monaural speaker separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee, B.-Y . Kim, and S. Watanabe, “TF-GridNet: Making time-frequency domain models great again for monaural speaker separation,” inProc. ICASSP, Rhodes, Greece, Jun. 2023

    2023

  9. [17]

    Meeting recognition with con- tinuous speech separation and transcription-supported di- arization,

    T. von Neumann, C. Boeddeker, T. Cord-Landwehr, M. Del- croix, and R. Haeb-Umbach, “Meeting recognition with con- tinuous speech separation and transcription-supported di- arization,” inProc. ICASSP Workshop on Hands-free Speech Communications and Microphone Arrays, Seoul, Kor...

    2024

  10. [18]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. Mcleavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inProc. ICML. Waikoloa, HI, USA: PMLR, Jul. 2023, pp. 28 492–28 518

    2023

  11. [19]

    Integration of speech separa- tion, diarization, and recognition for multi-speaker meetings: System description, comparison, and analysis,

    D. Raj, P. Denisov, Z. Chen, H. Erdogan, Z. Huang, M. He, S. Watanabe, J. Du, T. Yoshioka, Y . Luo, N. Kanda, J. Li, S. Wisdom, and J. R. Hershey, “Integration of speech separa- tion, diarization, and recognition for multi-speaker meetings: System description, comparison, and ...

    2021

  12. [20]

    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. Interspeech, 2024, pp. 1615–1619

    2024

  13. [21]

    DCF-DS: Deep cascade fusion of diarization and separation for speech recognition under realistic single- channel conditions,

    S.-T. Niu, J. Du, R.-Y . Wang, G.-B. Yang, T. Gao, J. Pan, and Y . Hu, “DCF-DS: Deep cascade fusion of diarization and separation for speech recognition under realistic single- channel conditions,” Preprint arXiv:2411.06667, 2024

    2024 arXiv

  14. [22]

    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,” inProc. CHiME Workshop, Dublin, Ireland, 2023, pp. 27–32

    2023

  15. [23]

    Lib- riSpeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- riSpeech: An ASR corpus based on public domain audio books,” inProc. ICASSP, Brisbane, Australia, Apr. 2015, pp. 5206–5210

    2015

  16. [24]

    SMS-WSJ: Database, performance measures, and baseline recipe for multi-channel source separation and recognition,

    L. Drude, J. Heitkaemper, C. Boeddeker, and R. Haeb- Umbach, “SMS-WSJ: Database, performance measures, and baseline recipe for multi-channel source separation and recognition,” Preprint arXiv:1910.13934, 2019

    1910 arXiv

  17. [25]

    Language mod- eling with deep transformers,

    K. Irie, A. Zeyer, R. Schlüter, and H. Ney, “Language mod- eling with deep transformers,” inProc. Interspeech, Graz, Austria, Sep. 2019, pp. 3905–3909

    2019

  18. [26]

    Con- former: 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, “Con- former: Convolution-augmented transformer for speech recognition,” inProc. Interspeech, Shanghai, China, Oct. 2020, pp. 5036–5040

    2020

  19. [27]

    WavLM: Large-scale self-supervised pre-training for full stack speech process- ing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “WavLM: Large-scale self-supervised pre-training for full stack speech process- ing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  20. [28]

    Performance measurement in blind audio source separation,

    E. Vincent, R. Gribonval, and C. Févotte, “Performance measurement in blind audio source separation,”IEEE trans- actions on audio, speech, and language processing, vol. 14, no. 4, pp. 1462–1469, 2006

    2006

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.