pith. sign in

arxiv: 1907.05584 · v1 · pith:K6Q4WAGAnew · submitted 2019-07-12 · 💻 cs.SD · cs.LG· eess.AS

Toeplitz Inverse Covariance based Robust Speaker Clustering for Naturalistic Audio Streams

Harishchandra Dubey , Abhijeet Sangwan , John Hansen This is my paper

Pith reviewed 2026-05-24 22:33 UTC · model grok-4.3

classification 💻 cs.SD cs.LGeess.AS
keywords speaker diarizationi-vectorMarkov random fieldToeplitz inverse covariancespeaker clusteringexpectation maximizationaudio streamsdiarization error rate
0
0 comments X

The pith

A Toeplitz inverse covariance matrix inside a Markov random field models speaker i-vector correlations to reduce diarization error rates.

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

The paper establishes that correlations among i-vectors from the same speaker can be represented by a Toeplitz-structured inverse covariance matrix in a Markov random field. This structure captures the sequential ordering of speaker turns in an audio stream. A variant of the expectation maximization algorithm that combines dynamic programming and the alternating direction method of multipliers yields a closed-form solution for the clustering. The resulting speaker clusters are evaluated after i-vector extraction, mean subtraction, PCA, and length normalization. Relative diarization error rate reductions of 43.22 percent on CRSS-PLTL, 29.37 percent on AMI IS1000a, and 9.21 percent on AMI IS1003b are reported against cosine K-means and movMF baselines.

Core claim

By representing each speaker's i-vector correlations as a Toeplitz inverse covariance matrix within a Markov random field, the method enables a closed-form solution for speaker clustering via a DP and ADMM variant of the EM algorithm, achieving relative DER reductions of 43.22% on CRSS-PLTL, 29.37% on AMI IS1000a, and 9.21% on AMI IS1003b.

What carries the argument

Toeplitz Inverse Covariance (TIC) matrix to represent the MRF correlation network for each speaker

If this is right

  • Speaker clustering can exploit the sequential structure of i-vectors belonging to the same speaker.
  • The DP+ADMM variant of EM supplies a closed-form update for the clustering parameters.
  • The four-step pipeline of ground-truth segmentation, i-vector extraction, post-processing, and TIC-MRF clustering produces measurable DER gains on naturalistic meeting data.
  • The model is directly compared against cosine K-means and movMF on the CRSS-PLTL and AMI corpora.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The structured covariance assumption may transfer to other sequential clustering tasks that involve ordered feature vectors.
  • Gains could change if i-vector extraction or post-processing steps are replaced by different front-ends.
  • The method might be tested in fully automatic segmentation settings rather than ground-truth segmentation.

Load-bearing premise

Correlations among i-vectors belonging to the same speaker can be adequately captured by a Toeplitz-structured inverse covariance matrix inside an MRF.

What would settle it

Running the proposed TIC-MRF clustering on the same datasets and observing no relative reduction or an increase in DER compared to the cosine K-means and movMF baselines.

Figures

Figures reproduced from arXiv: 1907.05584 by Abhijeet Sangwan, Harishchandra Dubey, John Hansen.

Figure 1
Figure 1. Figure 1: Block diagram of diarization pipeline employing proposed Toeplitz Inverse Covariance (TIC)-based speaker clustering. We perform mean subtraction in all experiments. Length-normalization is required only for cosine K-means and movMF [21] baselines. Algorithm 1 Assign-Clusters Input: GIVEN β ≥ 0, −LogL(i, j) = negative log-likelihood for i-th feature vector when it is assigned to j-th speaker cluster. K is t… view at source ↗
Figure 2
Figure 2. Figure 2: PLTL results: DER (%) for proposed and base￾line methods. No PCA, 21-PCA and 51-PCA represent cases where no dimensional reduction is performed, where 21 princi￾pal components and 51 principal components are chosen after PCA, respectively. Proposed approach achieves significant re￾duction in DER as compared to both baselines [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: AMI results: DER (%) for two meetings namely IS1000a and IS1003b. No PCA and 51-PCA represent cases where no dimension reduction is performed, and 51 princi￾pal components are chosen after PCA, respectively. Proposed speaker clustering gives less than 1% DER for both IS1000a and IS1003b (best case). The inverse covariance matrix, Θi is constrained to be block Toeplitz and λ is a nbXnb matrix so that it can… view at source ↗
read the original abstract

Speaker diarization determines who spoke and when? in an audio stream. In this study, we propose a model-based approach for robust speaker clustering using i-vectors. The ivectors extracted from different segments of same speaker are correlated. We model this correlation with a Markov Random Field (MRF) network. Leveraging the advancements in MRF modeling, we used Toeplitz Inverse Covariance (TIC) matrix to represent the MRF correlation network for each speaker. This approaches captures the sequential structure of i-vectors (or equivalent speaker turns) belonging to same speaker in an audio stream. A variant of standard Expectation Maximization (EM) algorithm is adopted for deriving closed-form solution using dynamic programming (DP) and the alternating direction method of multiplier (ADMM). Our diarization system has four steps: (1) ground-truth segmentation; (2) i-vector extraction; (3) post-processing (mean subtraction, principal component analysis, and length-normalization) ; and (4) proposed speaker clustering. We employ cosine K-means and movMF speaker clustering as baseline approaches. Our evaluation data is derived from: (i) CRSS-PLTL corpus, and (ii) two meetings subset of the AMI corpus. Relative reduction in diarization error rate (DER) for CRSS-PLTL corpus is 43.22% using the proposed advancements as compared to baseline. For AMI meetings IS1000a and IS1003b, relative DER reduction is 29.37% and 9.21%, respectively.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes a speaker clustering method for diarization that models correlations among i-vectors of the same speaker via a Markov Random Field whose precision matrix is constrained to be Toeplitz (TIC-MRF). Inference uses a dynamic-programming plus ADMM variant of EM claimed to admit a closed-form solution. The system pipeline consists of ground-truth segmentation, i-vector extraction, post-processing (mean subtraction, PCA, length normalization), and the proposed clustering step. On the CRSS-PLTL corpus the method yields a 43.22 % relative DER reduction versus cosine K-means and movMF baselines; on AMI meeting subsets IS1000a and IS1003b the reductions are 29.37 % and 9.21 %, respectively.

Significance. If the reported gains are shown to be statistically reliable and attributable to the TIC-MRF modeling choice rather than post-processing or baseline implementation details, the work would supply a concrete, structured way to exploit sequential dependence among speaker embeddings in diarization pipelines.

major comments (3)
  1. [Evaluation / Results] Evaluation (results tables and accompanying text): the abstract and results section report only point estimates of relative DER reduction; no statistical significance tests, bootstrap confidence intervals, or per-meeting variance are supplied, so it is impossible to judge whether the claimed 43.22 %, 29.37 % and 9.21 % improvements exceed sampling variability.
  2. [Proposed Method] Modeling section (TIC-MRF construction): the paper asserts that a Toeplitz-structured inverse covariance adequately captures same-speaker i-vector correlations, yet provides neither an empirical check (e.g., sample precision-matrix diagonals) nor an ablation that replaces the Toeplitz constraint with an unstructured or banded alternative; without such evidence the central modeling assumption remains unverified and the source of the reported gains cannot be isolated.
  3. [Algorithm] Algorithm section (DP+ADMM EM): the claim that the variant yields a closed-form optimum is stated without a derivation showing how the ADMM sub-problems preserve the claimed closed-form property or without convergence analysis; this detail is load-bearing for the assertion that the method is both tractable and superior to standard EM or the baselines.
minor comments (2)
  1. [System Pipeline] The post-processing pipeline (mean subtraction, PCA, length normalization) is applied identically to all methods; the manuscript should clarify whether any of these steps were tuned on the test data or whether they interact with the TIC-MRF objective.
  2. [Modeling] Notation for the MRF potential functions and the precise definition of the Toeplitz constraint (constant diagonals, bandwidth, etc.) should be stated explicitly with an equation reference.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and detailed review. The comments highlight important aspects of statistical rigor, modeling validation, and algorithmic transparency that we will address in the revision. Below we respond point-by-point to the major comments.

read point-by-point responses

  1. Referee: [Evaluation / Results] Evaluation (results tables and accompanying text): the abstract and results section report only point estimates of relative DER reduction; no statistical significance tests, bootstrap confidence intervals, or per-meeting variance are supplied, so it is impossible to judge whether the claimed 43.22 %, 29.37 % and 9.21 % improvements exceed sampling variability.

    Authors: We agree that reporting only point estimates limits the ability to assess reliability. In the revised manuscript we will add bootstrap confidence intervals (resampling over segments or meetings) and per-meeting DER breakdowns for both corpora. These additions will allow readers to evaluate whether the observed relative reductions exceed sampling variability. revision: yes

  2. Referee: [Proposed Method] Modeling section (TIC-MRF construction): the paper asserts that a Toeplitz-structured inverse covariance adequately captures same-speaker i-vector correlations, yet provides neither an empirical check (e.g., sample precision-matrix diagonals) nor an ablation that replaces the Toeplitz constraint with an unstructured or banded alternative; without such evidence the central modeling assumption remains unverified and the source of the reported gains cannot be isolated.

    Authors: The Toeplitz constraint is motivated by the stationary sequential dependence among i-vectors belonging to the same speaker, which is a natural modeling choice given the turn-based nature of the embeddings. While the original submission did not include an explicit empirical verification of the learned precision-matrix structure or an ablation against an unstructured MRF, the consistent gains over non-structured baselines (cosine K-means and movMF) provide indirect support. We will expand the modeling section to articulate this motivation more clearly and, space permitting, include a limited ablation comparing Toeplitz versus banded alternatives. revision: partial

  3. Referee: [Algorithm] Algorithm section (DP+ADMM EM): the claim that the variant yields a closed-form optimum is stated without a derivation showing how the ADMM sub-problems preserve the claimed closed-form property or without convergence analysis; this detail is load-bearing for the assertion that the method is both tractable and superior to standard EM or the baselines.

    Authors: The closed-form property follows from the combination of dynamic programming for the discrete assignment variables in the E-step and the ADMM solver for the constrained M-step, where the Toeplitz structure permits efficient closed-form updates for each sub-problem. We will add a dedicated appendix containing the full derivation of the ADMM sub-problems and a brief note on convergence (leveraging standard ADMM guarantees under convexity of the sub-problems). This will make the tractability claim fully transparent. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces the TIC-structured MRF for modeling i-vector correlations per speaker and a DP+ADMM variant of EM as modeling and algorithmic choices, then reports empirical relative DER reductions against cosine K-means and movMF baselines on three evaluation sets. No equations, self-citations, or steps are shown that reduce the claimed improvements to a fitted parameter defined by the same data, a self-referential definition, or a load-bearing self-citation chain. The central claim remains an independent modeling proposal whose validity is tested externally via diarization error rates.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract-only review; ledger populated from stated modeling premises only.

axioms (2)
  • domain assumption i-vectors extracted from segments of the same speaker are correlated
    Explicit premise used to justify the MRF network (abstract).
  • domain assumption A Toeplitz structure on the inverse covariance matrix is sufficient to represent the MRF correlation network for each speaker
    Core modeling choice enabling the closed-form solution (abstract).

pith-pipeline@v0.9.0 · 5826 in / 1368 out tokens · 25699 ms · 2026-05-24T22:33:24.992869+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 1 internal anchor

  1. [1]

    who spoke and when?

    Introduction Speaker diarization answers "who spoke and when?" in a multi- speaker audio stream [1]. Some of the practical applications of diarization technology include information retrieval [2], broad- cast news, meeting conversations, telephone calls, V oIP, digi- tal audio logging [3] and interaction analysis in Peer-Led Team Learning (PLTL) groups [4...

    work page 2019

  2. [2]

    Toeplitz Inverse Covariance based Robust Speaker Clustering for Naturalistic Audio Streams

    i-V ector Speaker Model Diarization involve extracting i-Vectors from short speech- segments (typically 1s) unlike speaker verification where com- plete utterance is used. Numerous techniques exist for cluster- ing i-Vectors using cosine distance [16]. The i-vector frame- work combines the speaker and channel variability sub-spaces of linear distortion mod...

    work page internal anchor Pith review Pith/arXiv arXiv 1907

  3. [3]

    Such temporal data has complicated structure where the underlying sequences of few fixed states repeat in definitive patterns

    Proposed Speaker Clustering Toeplitz Inverse Covariance (TIC)-based clustering was found to be suitable for segmenting the real-world time-series data such as fitness-tracker and driving data [22]. Such temporal data has complicated structure where the underlying sequences of few fixed states repeat in definitive patterns. Robust speaker clustering task poss...

    work page

  4. [4]

    Experiments, Results & Discussions In this study, we focus on speaker clustering and hence ground- truth segmentation information is adopted for all experiments to avoid errors from SAD and speaker segmentation steps (see Fig. 1). We conduct evaluations on: (i) CRSS-PLTL, and (ii) two meetings subset of AMI corpus, as detailed below. 4.1. CRSS-PLTL Eval S...

    work page

  5. [5]

    Such audio data are rich in noise, overlapped-speech and reverberation in addition to short conversational turns (1s)

    Summary & Conclusions In this paper, we leveraged the Toeplitz Inverse Covariance (TIC) estimation in speaker clustering task for naturalistic au- dio such as CRSS-PLTL corpus. Such audio data are rich in noise, overlapped-speech and reverberation in addition to short conversational turns (1s). The proposed approach accu- rately models the inner structure...

    work page

  6. [6]

    Speaker diarization: A review of recent research,

    X. Anguera, S. Bozonnet, N. Evans, C. Fredouille, G. Fried- land, and O. Vinyals, “Speaker diarization: A review of recent research,” IEEE Trans. on Audio, Speech, and Language Process- ing, vol. 20, no. 2, pp. 356–370, 2012

    work page 2012

  7. [7]

    Huijbregts, Segmentation, diarization and speech transcrip- tion: surprise data unraveled

    M. Huijbregts, Segmentation, diarization and speech transcrip- tion: surprise data unraveled. Citeseer, 2008

    work page 2008

  8. [8]

    Prof-Life-Log: Monitoring and assessment of human speech and acoustics using daily naturalistic audio streams,

    J. H. L. Hansen, A. Sangwan, A. Ziaei, H. Dubey, L. Kaushik, and C. Yu, “Prof-Life-Log: Monitoring and assessment of human speech and acoustics using daily naturalistic audio streams,” The Journal of the Acoustical Society of America, vol. 140, no. 4, pp. 3010–3010, 2016

    work page 2016

  9. [9]

    Incremental on-line clus- tering of speakers’ short segments,

    R. Aloni-Lavi, I. Opher, and I. Lapidot, “Incremental on-line clus- tering of speakers’ short segments,” in Proc. Odyssey 2018 The Speaker and Language Recognition Workshop , 2018, pp. 120– 127

    work page 2018

  10. [10]

    Large-scale speaker di- arization for long recordings and small collections,

    M. Huijbregts and D. A. van Leeuwen, “Large-scale speaker di- arization for long recordings and small collections,” IEEE Trans. on Audio, Speech, and Language Processing , vol. 20, no. 2, pp. 404–413, 2012

    work page 2012

  11. [11]

    Using speech tech- nology for quantifying behavioral characteristics in peer-led team learning sessions,

    H. Dubey, A. Sangwan, and J. H. L. Hansen, “Using speech tech- nology for quantifying behavioral characteristics in peer-led team learning sessions,” Computer Speech & Language , vol. 46, pp. 343–366, 2017

    work page 2017

  12. [12]

    A robust diarization system for measuring dominance in peer-led team learning groups,

    ——, “A robust diarization system for measuring dominance in peer-led team learning groups,” in IEEE Spoken Language Tech- nology Workshop (SLT), 2016, pp. 319–323

    work page 2016

  13. [13]

    Multi-stream audio analysis for knowledge extraction and understanding of small-group interactions in peer-led team learn- ing,

    J. H. L. Hansen, J. Alberte, N. Jones, H. Dubey, and A. Sang- wan, “Multi-stream audio analysis for knowledge extraction and understanding of small-group interactions in peer-led team learn- ing,” Seventh Annual Conference Peer-Led Team Learning Inter- national Society, the University of Texas at Dallas, Richardson, TX, USA, pp. 1–1, 2018

    work page 2018

  14. [14]

    CRSS-LDNN Long-duration naturalistic noise corpus containing multi-layer noise recordings for robust speech processing,

    J. H. L. Hansen, H. Dubey, and A. Sangwan, “CRSS-LDNN Long-duration naturalistic noise corpus containing multi-layer noise recordings for robust speech processing,” The Journal of the Acoustical Society of America, vol. 144, no. 3, pp. 1797–1797, 2018

    work page 2018

  15. [15]

    Leveraging Frequency-Dependent Kernel and DIP-based Clustering for Ro- bust Speech Activity Detection in Naturalistic Audio Streams,

    H. Dubey, A. Sangwan, and J. H. L. Hansen, “Leveraging Frequency-Dependent Kernel and DIP-based Clustering for Ro- bust Speech Activity Detection in Naturalistic Audio Streams,” IEEE/ACM Trans. on Audio, Speech and Language Processing , vol. 26, no. 11, pp. 2056–2071, 2018

    work page 2056

  16. [16]

    Robust feature clustering for unsupervised speech activity detection,

    ——, “Robust feature clustering for unsupervised speech activity detection,” in IEEE ICASSP, 2018, pp. 2726–2730

    work page 2018

  17. [17]

    Transfer learning using raw waveform sincnet for robust speaker diarization,

    ——, “Transfer learning using raw waveform sincnet for robust speaker diarization,” in IEEE ICASSP , Brighton, UK., 2019

    work page 2019

  18. [18]

    Front-end factor analysis for speaker verification,

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

    work page 2011

  19. [19]

    Speaker diariza- tion system for RT07 and RT09 meeting room audio,

    H. Sun, B. Ma, S. Z. K. Khine, and H. Li, “Speaker diariza- tion system for RT07 and RT09 meeting room audio,” in IEEE ICASSP, 2010, pp. 4982–4985

    work page 2010

  20. [20]

    Stream-based speaker segmentation using speaker factors and eigenvoices,

    F. Castaldo, D. Colibro, E. Dalmasso, P. Laface, and C. Vair, “Stream-based speaker segmentation using speaker factors and eigenvoices,” in IEEE ICASSP, 2008, pp. 4133–4136

    work page 2008

  21. [21]

    A study of the cosine distance-based mean shift for telephone speech diarization,

    M. Senoussaoui, P. Kenny, T. Stafylakis, and P. Dumouchel, “A study of the cosine distance-based mean shift for telephone speech diarization,” IEEE Trans. on Audio, Speech and Language Pro- cessing, vol. 22, no. 1, pp. 217–227, 2014

    work page 2014

  22. [22]

    Step-by-step and integrated approaches in broadcast news speaker diarization,

    S. Meignier, D. Moraru, C. Fredouille, J.-F. Bonastre, and L. Be- sacier, “Step-by-step and integrated approaches in broadcast news speaker diarization,” Computer Speech & Language, vol. 20, no. 2-3, pp. 303–330, 2006

    work page 2006

  23. [23]

    Recent Im- provements on ILP-based Clustering for Broadcast News Speaker Diarization,

    G. Dupuy, S. Meignier, P. Deléglise, and Y . Estéve, “Recent Im- provements on ILP-based Clustering for Broadcast News Speaker Diarization,” in ISCA Odyssey, 2014, pp. 187–193

    work page 2014

  24. [24]

    Generalized viterbi-based models for time-series segmentation and clustering applied to speaker diarization,

    I. Lapidot, A. Shoa, T. Furmanov, L. Aminov, A. Moyal, and J.-F. Bonastre, “Generalized viterbi-based models for time-series segmentation and clustering applied to speaker diarization,”Com- puter Speech & Language, vol. 45, pp. 1–20, 2017

    work page 2017

  25. [25]

    On the use of PLDA i-vector scoring for clustering short segments,

    I. Salmun, I. Opher, and I. Lapidot, “On the use of PLDA i-vector scoring for clustering short segments,” in Proc. Odyssey, 2016

    work page 2016

  26. [26]

    Robust speaker clustering using mixtures of von mises-fisher distributions for naturalistic audio streams,

    H. Dubey, A. Sangwan, and J. H. L. Hansen, “Robust speaker clustering using mixtures of von mises-fisher distributions for naturalistic audio streams,” in ISCA INTERSPEECH, 2018, pp. 3603–3607

    work page 2018

  27. [27]

    Toeplitz inverse covariance-based clustering of multivariate time series data,

    D. Hallac, S. Vare, S. Boyd, and J. Leskovec, “Toeplitz inverse covariance-based clustering of multivariate time series data,” in Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . ACM, 2017, pp. 215–223

    work page 2017

  28. [28]

    A speaker diarization system for studying peer-led team learning groups,

    H. Dubey, L. Kaushik, A. Sangwan, and J. H. L. Hansen, “A speaker diarization system for studying peer-led team learning groups,” in ISCA INTERSPEECH, 2016, pp. 2180–2184

    work page 2016

  29. [29]

    UTDallas-PLTL: Advancing multi-stream speech processing for interaction assessment in peer-led team learning,

    J. H. L. Hansen, H. Dubey, A. Sangwan, L. Kaushik, and V . Kothapally, “UTDallas-PLTL: Advancing multi-stream speech processing for interaction assessment in peer-led team learning,” The Journal of the Acoustical Society of America, vol. 143, no. 3, pp. 1869–1869, 2018

    work page 2018

  30. [30]

    The AMI meeting corpus,

    I. McCowan, J. Carletta, W. Kraaij, S. Ashby, S. Bourban, M. Flynn, M. Guillemot, T. Hain, J. Kadlec, V . Karaiskos et al., “The AMI meeting corpus,” in Proceedings of the 5th Interna- tional Conference on Methods and Techniques in Behavioral Re- search, vol. 88, 2005, p. 100

    work page 2005

  31. [31]

    Speaker recognition by machines and humans: A tutorial review,

    J. H. L. Hansen and T. Hasan, “Speaker recognition by machines and humans: A tutorial review,” IEEE Signal Processing Maga- zine, vol. 32, no. 6, pp. 74–99, 2015

    work page 2015

  32. [32]

    Eigenvoice model- ing with sparse training data,

    P. Kenny, G. Boulianne, and P. Dumouchel, “Eigenvoice model- ing with sparse training data,” IEEE Trans. on speech and audio processing, vol. 13, no. 3, pp. 345–354, 2005

    work page 2005

  33. [33]

    S. L. Lauritzen, Graphical models . Clarendon Press, 1996, vol. 17

    work page 1996

  34. [34]

    Assessing the Impact of a Multi-Disciplinary Peer-Led- Team Learning Program on Undergraduate STEM Education,

    K. Carlson, D. Turvold Celotta, E. Curran, M. Marcus, and M. Loe, “Assessing the Impact of a Multi-Disciplinary Peer-Led- Team Learning Program on Undergraduate STEM Education,” Journal of University Teaching & Learning Practice , vol. 13, no. 1, p. 5, 2016

    work page 2016

  35. [35]

    Studying the relationship between physical and language environments of children: Who’s speaking to whom and where?

    A. Sangwan, J. H. L. Hansen, D. W. Irvin, S. Crutchfield, and C. R. Greenwood, “Studying the relationship between physical and language environments of children: Who’s speaking to whom and where?” in IEEE Signal Proc.Education Workshop, Salt Lake City, Utah, 2015, pp. 49–54

    work page 2015

  36. [36]

    Fast speaker diarization using a high-level scripting language,

    E. Gonina, G. Friedland, H. Cook, and K. Keutzer, “Fast speaker diarization using a high-level scripting language,” inIEEE Workshop on Automatic Speech Recognition and Understanding (ASRU), 2011, pp. 553–558

    work page 2011

  37. [37]

    Efficient online spherical k-means clustering,

    S. Zhong, “Efficient online spherical k-means clustering,” in IEEE International Joint Conference on Neural Networks, vol. 5, 2005, pp. 3180–3185

    work page 2005

  38. [38]

    Analysis of i-vector length normalization in speaker recognition systems,

    D. Garcia-Romero and C. Y . Espy-Wilson, “Analysis of i-vector length normalization in speaker recognition systems,” inISCA IN- TERSPEECH, 2011, pp. 249–252

    work page 2011

  39. [39]

    NIST DER script for RT evaluations,

    “NIST DER script for RT evaluations,” (Date last accessed 9- Jan-2018). [Online]. Available: as part of the Speech Recognition Scoring Toolkit (SCTK): ftp://jaguar.ncsl.nist.gov/pub/sctk-2.4. 10-20151007-1312Z.tar.bz2

    work page 2018