Pith. sign in

REVIEW 4 major objections 9 minor 1 cited by

Speaker recognition network found to hierarchically organize voices by gender then nationality

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-04 14:44 UTC pith:7ECUX2WT

load-bearing objection HCCM extends cluster-class matching to hierarchical clusters and applies it to speaker recognition embeddings — a real methodological contribution, but the central dendrogram interpretation rests on single-linkage clustering without addressing chaining. the 4 major comments →

arxiv 2604.23354 v3 pith:7ECUX2WT submitted 2026-04-25 eess.AS cs.AIeess.SP

Explainable AI in Speaker Recognition -- Making Latent Representations Understandable

Yanze Xu , Wenwu Wang , Mark D. Plumbley This is my paper
classification eess.AS cs.AIeess.SP
keywords speaker recognitionexplainable AIhierarchical clusteringrepresentation spacecluster-class matchingsingle-linkage clusteringHDBSCANLiebig score
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the internal representation space of a trained speaker recognition network is not flat but hierarchically structured. By applying single-linkage clustering to embeddings extracted from a ResNet34-based speaker recognition model, the authors find that the top level of the resulting hierarchy cleanly separates male from female speakers (with near-perfect matching), while lower levels branch into conjunctive categories combining gender and nationality, such as UK&male or Ireland&female. To interpret these clusters, the authors design Hierarchical Cluster-Class Matching (HCCM), an algorithm that pairs each hierarchical cluster with the semantic class (individual or conjunctive) that best explains it. They also introduce the Liebig score (L-score), a metric that replaces the F-score and reports matching quality as the single limiting factor (precision or recall), making the degree of match directly interpretable. The central object is the annotated dendrogram: a tree of representation clusters where gender sits at the top and gender-by-nationality conjunctions sit below, suggesting the network has implicitly learned a structured taxonomy of voice characteristics.

Core claim

The paper's central finding is that single-linkage clustering applied to 4-second speaker embeddings from a well-trained speaker recognition network produces a hierarchy whose upper levels match gender classes with near-1.0 cluster-class matching scores, and whose lower levels match conjunctive gender-and-nationality classes (e.g., USA&male, UK&female). The authors also show that HDBSCAN (which adds density constraints) consistently produces lower matching scores than unconstrained SLINK, and that shorter audio clips degrade the hierarchy's semantic alignment. The HCCM method, combined with the L-score, provides a systematic way to label each cluster with its best-matched semantic class and,

What carries the argument

Single-Linkage Clustering (SLINK), Hierarchical Cluster-Class Matching (HCCM), Liebig score (L-score)

Load-bearing premise

The paper assumes that single-linkage clustering without any density constraint (minPts=0) produces the most semantically meaningful hierarchy because it achieves the highest cluster-class matching scores, but does not address that single-linkage is known to produce chaining artifacts where clusters merge through thin bridges of points rather than genuine structural connections.

What would settle it

Apply a clustering algorithm robust to chaining (e.g., average-linkage or Ward's method) to the same embeddings; if the gender-first, nationality-second hierarchy disappears or degrades substantially, the structure may be an artifact of SLINK's sensitivity to noise bridges rather than a genuine property of the representation space.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The speaker recognition network implicitly learns a structured taxonomy of human voices, separating gender first and nationality-gender conjunctions second, without being explicitly trained on either attribute.
  • The Liebig score offers a replacement for F-score in any cluster-class matching task where knowing the limiting factor (precision vs. recall) is more informative than a harmonic mean.
  • Longer audio inputs produce more semantically interpretable internal hierarchies, linking model generalization quality directly to the interpretability of its representation space.
  • The observation that Ireland, UK, and Canada male voice clusters merge at a higher level suggests the network captures cross-national acoustic similarities, potentially reflecting shared linguistic or historical ties.
  • The HCCM framework is generalizable: any neural network's representation space can be probed for hierarchical structure by clustering embeddings and matching clusters against available metadata labels.

Where Pith is reading between the lines

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

  • If the network organizes representations hierarchically by gender and nationality without being trained on these labels, then the representation space may encode phylogenetic or sociolinguistic structure in voice that parallels human perceptual categories.
  • The fact that SLINK without density constraints outperforms HDBSCAN could mean the representation space is genuinely tree-structured rather than density-modulated, which would have implications for how speaker diarization systems should partition embeddings.
  • If the hierarchy generalizes to other speaker recognition architectures, one could use the depth and branching factor of the dendrogram as a diagnostic for how much paralinguistic information a given model encodes, beyond its primary task of speaker identity.
  • The gender-first structure raises the question of whether this reflects an acoustic reality (formant differences) or a dataset bias (unequal representation of genders across nationalities in VoxCeleb).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 9 minor

Summary. The manuscript proposes a framework for analyzing, visualizing, and interpreting hierarchical clustering structure within the latent representation space of a pre-trained speaker recognition network. Two hierarchical clustering algorithms (SLINK and HDBSCAN) are applied to speaker embeddings extracted from a ResNet34 model trained on VoxCeleb2. The quality of the resulting hierarchical clusterings is evaluated using the Cluster-Class Matching (CCM) method of Rosenberg and Hirschberg. The best-performing configuration (SLINK on 4-second embeddings) is then visualized as a dendrogram and semantically interpreted via a proposed Hierarchical Cluster-Class Matching (HCCM) algorithm, which greedily matches clusters to predefined semantic classes (gender, nationality, and their conjunctions). A new metric, the Liebig score (L-score), is introduced to replace the F-score in quantifying individual cluster-class matching quality, motivated by Liebig's law of the minimum. The central finding is that the dendrogram exhibits a semantically meaningful hierarchy: gender splits at the top, and gender-nationality conjunctions (e.g., UK&male, Ireland&female) appear at lower levels.

Significance. The paper addresses a genuine gap in the XAI literature for speaker recognition: while flat clustering of network embeddings has been studied, hierarchical clustering structure and its semantic interpretation have received little attention. The HCCM method and L-score metric are concrete, falsifiable methodological contributions. The L-score in particular offers a principled alternative to the F-score by directly identifying the limiting factor (precision or recall) in each cluster-class match, which is a useful diagnostic tool. The dendrogram annotations in Fig. 8 provide a tangible, checkable visualization of the claimed hierarchical semantic structure. The use of a publicly available pre-trained network and standard clustering implementations supports reproducibility, though no code or data release is provided by the authors. The central claim—that the representation space encodes a hierarchy from gender to gender-nationality conjunctions—is interesting and, if robust, would be a meaningful finding for the interpretability of speaker recognition systems.

major comments (4)
  1. §VI-A, Fig. 7 and §VI-B, Fig. 8: The central claim—that the dendrogram reveals a genuine hierarchical organization (gender at top, nationality-gender conjunctions below)—rests entirely on SLINK with minPts=0. Single-linkage clustering is well-known to be susceptible to chaining effects, where clusters merge via thin bridges of points, potentially distorting the merge order that the semantic interpretation depends on. The CCM scores used to select this configuration measure how well flat clusters at various cut levels align with predefined classes, but they do not validate the merge order that produces the hierarchy. The paper does not discuss this known limitation or test whether the observed hierarchy is robust to the clustering algorithm choice. A straightforward robustness check would be to examine whether HDBSCAN's dendrogram (at a moderate minPts) shows the same hierarchical pattern
  2. §V (Experimental Procedures): No code, data, or configuration details sufficient for reproduction are provided. The paper uses a pre-trained network from Chung et al. [11,54], the HDBSCAN implementation from McInnes et al. [53], and VoxCeleb1 test set annotations, but the exact extraction protocol (e.g., which specific utterances, how many embeddings per speaker, whether embeddings are L2-normalized, the distance metric used for SLINK) is not fully specified. The dendrogram in Fig. 8 is also truncated to clusters with >800 representations, and the L-score display threshold of 0.25 is ad hoc. Without reproducible code or a more detailed experimental protocol, the results cannot be independently verified. This should be addressed before publication.
  3. §IV-B, Eq. (4): The HCCM greedy matching procedure removes the matched semantic class c*_l from C^(l-1) at each iteration but does not remove the matched cluster k*_l from K. This means the same cluster can be matched to multiple semantic classes (as acknowledged in §VI-B for the France&female/Norway&female case). While this is a defensible design choice, it has a consequence for the interpretation: if a single cluster is the best match for multiple classes, the 'hierarchical' interpretation becomes ambiguous. The paper should explicitly discuss this design decision, its implications for the claimed hierarchy, and whether enforcing one-to-one matching (or reporting results under both schemes) would change the dendrogram annotations in Fig. 8.
  4. §IV-C, Eq. (5): The L-score is defined as min(precision, recall) = |c∩k|/max(|c|,|k|). While the motivation from Liebig's law is reasonable, the L-score has a known limitation: it can be zero whenever either precision or recall is zero, even if the other is high. More importantly, the L-score does not account for chance agreement. For the dendrogram annotations in Fig. 8, some L-scores are as low as 0.26 (pre:0.26 for one cluster), yet these are still displayed and interpreted. The paper should discuss whether there is a principled threshold below which a match should not be interpreted, and whether the L-score's behavior at low values is well-calibrated for the clustering application. The claim that L-score is 'more interpretable' than F-score should also be scoped more carefully.
minor comments (9)
  1. Algorithm 1 caption: 'Find a flat cluster by DBSCAN (asdfasdfasdfasdfasdfasdf)' contains placeholder text that should be corrected.
  2. §III-A, Fig. 4 caption: 'Pesudocode' should be 'Pseudocode'.
  3. §IV-A, Algorithm 1, line 6: The condition 'if label(q) ≠ undefined then' has an empty body (lines 7-8 are blank). This appears to be a formatting error; the algorithm should presumably skip already-labeled points.
  4. §IV-A: The claim that 'the equivalence between HDBSCAN and MST-based SLINK on the mutual reachability distance has not been explicitly discussed in prior work' is a strong novelty claim. Campello et al. [25] and McInnes et al. [53] discuss the MST-based formulation of HDBSCAN in some detail. The contribution here should be clarified: is it a new proof, or a new exposition of the relationship in the context of network representations?
  5. §VI-B, Fig. 8: The dendrogram is very dense and the text labels are small. Consider providing a zoomed-in version of key subtrees (e.g., the male subtree and female subtree separately) to improve readability.
  6. §VI-B: The speculative interpretations (e.g., 'historical and linguistic ties among Ireland, the UK, and Canada' explaining cluster merging; 'historical ties between Norway and France' explaining female voice similarity) are presented without supporting evidence. These should be clearly labeled as hypotheses rather than findings.
  7. §VII (Discussion): This section is very brief and only discusses the audio-length effect. A more thorough discussion of limitations (chaining, ad hoc thresholds, single network, single dataset) and threats to validity would strengthen the paper.
  8. §II-C: The related work on hierarchical clustering for speaker diarization [43,44,45] is mentioned but not discussed in relation to the current work. A sentence clarifying how this paper's goals (XAI/interpretability) differ from diarization (performance) would help.
  9. References: Some entries appear to have formatting issues (e.g., Ref [50] 'Introduction to information retrieval. Syngress Publishing,, 2008' has double commas; Ref [13] has 'vol. 356, p. 360' which seems incomplete; verify accuracy).

Simulated Author's Rebuttal

4 responses · 1 unresolved

We thank the referee for a careful and constructive report. The referee raises four major points concerning: (1) robustness of the claimed hierarchy to single-linkage chaining effects, (2) reproducibility of experimental details, (3) the design and implications of allowing one cluster to match multiple semantic classes in HCCM, and (4) the calibration and thresholding of the L-score. We agree that all four points warrant attention and will revise the manuscript accordingly. For point 1, we will add a discussion of chaining effects and conduct a robustness check using HDBSCAN's dendrogram at moderate minPts values. For point 2, we will expand the experimental protocol with full extraction and clustering details and commit to releasing code. For point 3, we will explicitly discuss the one-to-many matching design, its implications, and report results under a one-to-one matching constraint. For point 4, we will add discussion of L-score behavior at low values, principled thresholds, and scope the 'more interpretable' claim. We cannot fully resolve whether the observed hierarchy is entirely free of chaining artifacts, as this is an inherent limitation of single-linkage clustering, but we will transparently address this.

read point-by-point responses

  1. Referee: §VI-A, Fig. 7 and §VI-B, Fig. 8: The central claim rests entirely on SLINK with minPts=0. Single-linkage clustering is susceptible to chaining effects. CCM scores do not validate the merge order. The paper does not discuss this limitation or test robustness to algorithm choice. A robustness check using HDBSCAN's dendrogram at moderate minPts would address this.

    Authors: The referee is correct that single-linkage clustering is susceptible to chaining effects and that CCM scores evaluate flat-cluster alignment at cut levels rather than validating the merge order itself. We will revise the manuscript to explicitly acknowledge this limitation. Regarding the robustness check: the equivalence between HDBSCAN and MST-based SLINK on mutual reachability distance, which we establish in Section IV-A, means that HDBSCAN at moderate minPts modifies the distance metric (via the mutual reachability distance) but preserves the same hierarchical structure. We will examine whether HDBSCAN's dendrogram at moderate minPts (e.g., minPts=2 or 4) shows the same gender-at-top, nationality-gender-conjunctions-below pattern and report the results. We note that at minPts>0, the CCM scores in Fig. 7 do decline, which may reflect either the density constraint removing valid cluster members or genuine degradation of the hierarchy. We will discuss both possibilities. We acknowledge that we cannot fully rule out that chaining contributes to the observed merge order; this is an inherent limitation of single-linkage that we will state transparently. revision: partial

  2. Referee: §V: No code, data, or configuration details sufficient for reproduction are provided. The exact extraction protocol, L2-normalization, distance metric for SLINK, dendrogram truncation threshold, and L-score display threshold are not fully specified.

    Authors: We agree that the experimental protocol should be more fully specified. In the revision we will add the following details: (1) embeddings are extracted from the penultimate layer of the ResNet34 model for each audio clip duration (0.2, 1, 2, 4 seconds), with one embedding per utterance; (2) embeddings are L2-normalized; (3) the distance metric for SLINK is Euclidean distance (as stated in Section V, the HDBSCAN implementation uses Euclidean-derived mutual reachability distance, and minPts=0 reduces this to Euclidean); (4) the dendrogram in Fig. 8 is truncated to clusters with >800 representations for visual clarity, and we will state this threshold's rationale; (5) the L-score display threshold of 0.25 was chosen to include matches with at least 25% overlap while keeping the figure readable, and we will discuss this choice. We will also commit to releasing code and configuration files upon acceptance to support reproducibility. revision: yes

  3. Referee: §IV-B, Eq. (4): HCCM removes the matched class but not the matched cluster from K, allowing one cluster to match multiple classes. This makes the hierarchical interpretation ambiguous. The paper should discuss this design decision, its implications, and whether one-to-one matching would change Fig. 8.

    Authors: This is a fair point. The decision to remove only the matched class (not the matched cluster) was deliberate: it allows a single cluster to be interpreted as a union of multiple semantic classes when the network genuinely conflates them, as in the France&female/Norway&female case discussed in Section VI-B. However, we agree that this design choice and its implications for the hierarchical interpretation should be discussed explicitly. In the revision we will: (1) add a paragraph explaining the rationale for allowing one-to-many cluster-to-class matching; (2) discuss the ambiguity it introduces; (3) report the dendrogram annotations under a one-to-one matching constraint (where each cluster is also removed after matching) and compare the results. If the one-to-one scheme changes any annotations in Fig. 8, we will report which ones. revision: yes

  4. Referee: §IV-C, Eq. (5): The L-score can be zero when either precision or recall is zero. It does not account for chance agreement. Some L-scores as low as 0.26 are displayed and interpreted. The paper should discuss a principled threshold and scope the 'more interpretable' claim.

    Authors: The referee raises valid concerns about L-score calibration at low values. We will address these as follows. First, we will add discussion of the L-score's behavior at low values, noting that an L-score of 0.26 means the match is limited by whichever of precision or recall equals 0.26, and that such a low score indicates the cluster is a weak match that should be interpreted cautiously. Second, regarding a principled threshold: we will discuss the possibility of using a chance-corrected baseline (e.g., comparing the L-score to the expected overlap under random assignment) and will report whether the 0.25 threshold is above chance level for the cluster and class sizes involved. Third, we will scope the claim that L-score is 'more interpretable' than F-score: the L-score's advantage is specifically that it identifies the limiting factor (precision or recall) and its value directly equals that factor, whereas the F-score is a harmonic mean without a direct proportional interpretation. We will clarify that this does not mean the L-score is superior in all respects—it is more conservative and can be zero when one factor is zero—and will state these trade-offs explicitly. revision: yes

standing simulated objections not resolved
  • The fundamental question of whether the observed merge order in the SLINK dendrogram is entirely free of chaining artifacts cannot be definitively resolved, as single-linkage clustering is inherently susceptible to this effect. While we can examine robustness via HDBSCAN and discuss the limitation transparently, we cannot fully eliminate the possibility that chaining influences the hierarchy.

Circularity Check

0 steps flagged

No significant circularity found; derivation is self-contained against external benchmarks.

full rationale

The paper's derivation chain is self-contained. The central claim—that SLINK on 4-second speaker embeddings reveals a hierarchy with gender at the top and nationality-gender conjunctions below—depends on three independent inputs: (1) a pre-trained speaker recognition network from Chung et al. [11], (2) standard hierarchical clustering algorithms (SLINK, HDBSCAN) applied unsupervised to embeddings, and (3) VoxCeleb1 test-set metadata labels used for semantic class definitions. The CCM evaluation method is from Rosenberg et al. [26] (external, not self-cited). The HCCM method (Eq. 4) performs greedy pairwise matching between unsupervised clusters and predefined label divisions—these are independent sources, so the matching is not circular by construction. The L-score (Eq. 5) is a new metric defined as min(precision, recall), which does not reduce to any fitted parameter. The SLINK/HDBSCAN equivalence argument (Section IV-A) is a mathematical comparison of algorithm pseudocodes, not a self-citation chain. The only self-citation is [61] (Xu, Wang et al.), which appears solely in the conclusion's future-work paragraph about inviting human experts and is not load-bearing for any claim in the paper. No step in the derivation reduces to its inputs by definition or by self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities, particles, or forces are introduced. The paper proposes algorithmic constructs (HCCM, L-score) but these are methods, not postulated entities.

free parameters (4)
  • minPts = swept over {0, 2, 4, 6, 8, 12, 16, 21, 27}
    Density parameter for HDBSCAN/SLINK; minPts=0 (SLINK) selected as optimal based on CCM scores
  • audio_length = selected 4 seconds as optimal
    Four input lengths tested (0.2, 1, 2, 4 seconds); 4-second chosen based on CCM scores
  • L-score threshold for display = 0.25
    Minimum L-score for annotating cluster-class pairs on dendrogram; chosen by hand for visual clarity
  • minimum cluster size for dendrogram display = 800
    Threshold for retaining clusters in the visualized dendrogram; chosen by hand for conciseness
axioms (4)
  • domain assumption Neural networks organize representations in ways analogous to human knowledge organization
    Stated in Section I: 'we assume that the neural network has a similar ability to organise representations of different inputs.' This motivates the entire analysis.
  • domain assumption Higher CCM matching degree implies more semantically meaningful clustering
    Section III-B and VI-A: the paper assumes that alignment with predefined grouping information reflects semantic meaningfulness of the clustering phenomenon.
  • ad hoc to paper Liebig's law of the minimum applies to cluster-class matching performance
    Section IV-C: 'we assume that the matching performance between a representation cluster and a predefined representation class division is primarily limited by the weakest matching factor.' This justifies using min(precision, recall) as the L-score.
  • domain assumption Predefined semantic classes (gender, nationality) are appropriate dimensions for interpreting speaker recognition representations
    Section V: the paper uses VoxCeleb1 annotations for gender and nationality as the semantic classes for HCCM interpretation, assuming these are the relevant interpretive dimensions.

pith-pipeline@v1.1.0-glm · 21496 in / 3490 out tokens · 77334 ms · 2026-07-04T14:44:13.529465+00:00 · methodology

0 comments
read the original abstract

Neural networks can be trained to learn task-relevant representations from data. Understanding how these networks make decisions falls within the Explainable AI (XAI) domain. This paper proposes to study an XAI topic: analysing, visualising and understanding the unknown organisation of network representations, particularly those a speaker recognition network learns from utterances, for recognising speaker identity. Past studies have employed algorithms (e.g. K-means) to analyse the different ways in which network representations can be naturally grouped into clusters, i.e. to analyse different flat clustering phenomena within the space defined by those representations. In contrast, this work applies two algorithms -- Single-Linkage Clustering (SLINK) and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) -- to analyse the different ways in which representations from the speaker recognition network can form clusters with hierarchical relationships, i.e., to analyse different hierarchical clustering phenomena within the representation space of the speaker recognition network. Furthermore, an algorithm called Hierarchical Cluster-Class Matching (HCCM) is designed to semantically interpret one of the above hierarchical clustering phenomena analysed using SLINK. Given the clusters representing this phenomenon, HCCM identifies which ones best match individual semantic classes related to gender and nationality (e.g.\ male, female, Ireland, UK) and and-logic conjunctions of these classes (e.g.\ female and Ireland). The Liebig score metric is also proposed within HCCM to quantify the matching quality of each cluster-class pair and diagnose the factor that limits each match.

Figures

Figures reproduced from arXiv: 2604.23354 by Mark D. Plumbley, Wenwu Wang, Yanze Xu.

Figure 1
Figure 1. Figure 1: An approximate 2-dimensional visualisation for the view at source ↗
Figure 2
Figure 2. Figure 2: An illustration of the Hierarchical Cluster-Class Match view at source ↗
Figure 3
Figure 3. Figure 3: An illustration for interpreting the matching degree view at source ↗
Figure 4
Figure 4. Figure 4: Pesudocode of the DBSCAN from RJGB Campello’s view at source ↗
Figure 5
Figure 5. Figure 5: An illustration of intersecting predefined representation view at source ↗
Figure 6
Figure 6. Figure 6: An overview of experimental procedures separately, because when the density constraint parameter minP ts in McInnes et al.’s HDBSCAN implementation is set to 0, the mutual reachability distance space coincides with the original Euclidean distance space, in which case McInnes et al.’s HDBSCAN implementation reduces to running SLINK directly in the original Euclidean distance space. More detailed setups of h… view at source ↗
Figure 7
Figure 7. Figure 7: Cluster-class matching results [26] for evaluating the hierarchical clustering results obtained by applying SLINK (i.e. minP ts = 0) and HDBSCAN (i.e. minP ts = 2, 4, 6, 8, 12, 16, 21, 27)) to representations (i.e. embeddings) of 0.2, 1, 2, and 4-sec audios. hierarchical representation clusters of both Fig. 7b and Fig. 7c are consistently produced by applying SLINK to 4-second audio representations, achiev… view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Explainable AI in Speaker Recognition -- Attention Map Visualisation and Evaluation

    eess.AS 2026-06 unverdicted novelty 5.0

    The paper introduces Modified RISE-eval to evaluate GradCAM and LayerCAM attention maps on speaker recognition networks and reports distinct advantages for each method under different conditions.

Reference graph

Works this paper leans on

61 extracted references · 61 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    work page 2016

  2. [2]

    Deep learning,

    Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,”nature, vol. 521, no. 7553, pp. 436–444, 2015

    work page 2015

  3. [3]

    DARPA’s explainable artificial intelligence (XAI) program,

    D. Gunning and D. Aha, “DARPA’s explainable artificial intelligence (XAI) program,”AI magazine, vol. 40, no. 2, pp. 44–58, 2019

    work page 2019

  4. [4]

    Explain- able AI: A brief survey on history, research areas, approaches and challenges,

    F. Xu, H. Uszkoreit, Y . Du, W. Fan, D. Zhao, and J. Zhu, “Explain- able AI: A brief survey on history, research areas, approaches and challenges,” inNatural Language Processing and Chinese Computing: 8th CCF International Conference, NLPCC 2019, Dunhuang, China, October 9–14, 2019, Proceedings, Part II 8. Springer, 2019, pp. 563– 574

    work page 2019

  5. [5]

    Explainable AI: A review of machine learning interpretability methods,

    P. Linardatos, V . Papastefanopoulos, and S. Kotsiantis, “Explainable AI: A review of machine learning interpretability methods,”Entropy, vol. 23, no. 1, p. 18, 2020

    work page 2020

  6. [6]

    An overview of the supervised machine learning methods,

    V . Nasteski, “An overview of the supervised machine learning methods,” Horizons. b, vol. 4, no. 51-62, p. 56, 2017

    work page 2017

  7. [7]

    How humans learn and represent networks,

    C. W. Lynn and D. S. Bassett, “How humans learn and represent networks,”Proceedings of the National Academy of Sciences, vol. 117, no. 47, pp. 29 407–29 415, 2020

    work page 2020

  8. [8]

    Selective attention

    W. A. Johnston and V . J. Dark, “Selective attention.”Annual Review of Psychology, 1986

    work page 1986

  9. [9]

    Exploring the Encoding Layer and Loss Function in End-to-End Speaker and Language Recognition System

    W. Cai, J. Chen, and M. Li, “Exploring the encoding layer and loss function in end-to-end speaker and language recognition system,”arXiv preprint arXiv:1804.05160, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  10. [10]

    V oxceleb: A large-scale speaker identification dataset,

    A. Nagrani, J. S. Chung, and A. Zisserman, “V oxceleb: A large-scale speaker identification dataset,”arXiv preprint arXiv:1706.08612, 2017

    work page arXiv 2017

  11. [11]

    In defence of metric learning for speaker recognition,

    J. S. Chung, J. Huh, S. Mun, M. Lee, H. S. Heo, S. Choe, C. Ham, S. Jung, B.-J. Lee, and I. Han, “In defence of metric learning for speaker recognition,”arXiv preprint arXiv:2003.11982, 2020

    work page arXiv 2003

  12. [12]

    Efficient integrated features based on pre-trained models for speaker verification,

    Y . Li, W. Guan, H. Huang, S. Miao, Q. Su, L. Li, and Q. Hong, “Efficient integrated features based on pre-trained models for speaker verification,” inProc. Interspeech 2024, 2024, pp. 2140–2144

    work page 2024

  13. [13]

    Flat clustering,

    C. D. Manning, P. Raghavan, and H. Sch ¨utze, “Flat clustering,”Intro- duction to Information Retrieval, vol. 356, p. 360, 2008

    work page 2008

  14. [14]

    Some methods for classification and analysis of multivariate observations,

    J. MacQueenet al., “Some methods for classification and analysis of multivariate observations,” inProceedings of the fifth Berkeley Sym- posium on Mathematical Statistics and Probability, vol. 1, no. 14. Oakland, CA, USA, 1967, pp. 281–297

    work page 1967

  15. [15]

    Investigating deep neural network internal clustering and generalization properties,

    G. Peiffer, “Investigating deep neural network internal clustering and generalization properties,” Master’s thesis, Universit ´e catholique de Louvain, Louvain, Belgium, 2021

    work page 2021

  16. [16]

    Intraclass clustering: An implicit learning ability that regularizes DNNs,

    S. Carbonnelle and C. De Vleeschouwer, “Intraclass clustering: An implicit learning ability that regularizes DNNs,” inInternational Con- ference on Learning Representations, 2020

    work page 2020

  17. [17]

    Deep clustering for unsupervised learning of visual features,

    M. Caron, P. Bojanowski, A. Joulin, and M. Douze, “Deep clustering for unsupervised learning of visual features,” inProceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 132–149

    work page 2018

  18. [18]

    The tree of diffusion life: Evolutionary embeddings to understand the generation process of diffusion models,

    V . Prasad, H. van Gorp, C. Humer, A. Vilanova, and N. Pezzotti, “The tree of diffusion life: Evolutionary embeddings to understand the generation process of diffusion models,”arXiv e-prints, pp. arXiv–2406, 2024

    work page 2024

  19. [19]

    Frame- level phoneme-invariant speaker embedding for text-independent speaker recognition on extremely short utterances,

    N. Tawara, A. Ogawa, T. Iwata, M. Delcroix, and T. Ogawa, “Frame- level phoneme-invariant speaker embedding for text-independent speaker recognition on extremely short utterances,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 6799–6803. 14

    work page 2020

  20. [20]

    Visualizing data using t-SNE

    L. Van der Maaten and G. Hinton, “Visualizing data using t-SNE.” Journal of Machine Learning Research, vol. 9, no. 11, 2008

    work page 2008

  21. [21]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    L. McInnes, J. Healy, and J. Melville, “Umap: Uniform manifold approximation and projection for dimension reduction,”arXiv preprint arXiv:1802.03426, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  22. [22]

    Comprehensive survey on hierarchical clustering algorithms and the recent developments,

    X. Ran, Y . Xi, Y . Lu, X. Wang, and Z. Lu, “Comprehensive survey on hierarchical clustering algorithms and the recent developments,” Artificial Intelligence Review, vol. 56, no. 8, pp. 8219–8264, 2023

    work page 2023

  23. [23]

    Minimum spanning trees and single linkage cluster analysis,

    J. C. Gower and G. J. Ross, “Minimum spanning trees and single linkage cluster analysis,”Journal of the Royal Statistical Society: Series C (Applied Statistics), vol. 18, no. 1, pp. 54–64, 1969

    work page 1969

  24. [24]

    SLINK: an optimally efficient algorithm for the single-link cluster method,

    R. Sibson, “SLINK: an optimally efficient algorithm for the single-link cluster method,”The Computer Journal, vol. 16, no. 1, pp. 30–34, 1973

    work page 1973

  25. [25]

    Density-based clustering based on hierarchical density estimates,

    R. J. G. B. Campello, D. Moulavi, and J. Sander, “Density-based clustering based on hierarchical density estimates,” inAdvances in Knowledge Discovery and Data Mining, J. Pei, V . S. Tseng, L. Cao, H. Motoda, and G. Xu, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2013, pp. 160–172

    work page 2013

  26. [26]

    V-measure: A conditional entropy- based external cluster evaluation measure,

    A. Rosenberg and J. Hirschberg, “V-measure: A conditional entropy- based external cluster evaluation measure,” inProceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Process- ing and Computational Natural Language Learning (EMNLP-CoNLL), 2007, pp. 410–420

    work page 2007

  27. [27]

    A statistical method for evaluating systematic relationships,

    R. R. Sokal, C. D. Micheneret al., “A statistical method for evaluating systematic relationships,” 1958

    work page 1958

  28. [28]

    Morzycki,Modification

    M. Morzycki,Modification. Cambridge University Press, 2016

    work page 2016

  29. [29]

    V on Liebig’s law of the minimum and plankton ecology (1899–1991),

    H. De Baar, “V on Liebig’s law of the minimum and plankton ecology (1899–1991),”Progress in Oceanography, vol. 33, no. 4, pp. 347–386, 1994

    work page 1991

  30. [30]

    A review of the f-measure: Its history, properties, criticism, and alternatives,

    P. Christen, D. J. Hand, and N. Kirielle, “A review of the f-measure: Its history, properties, criticism, and alternatives,”ACM Computing Surveys, vol. 56, no. 3, pp. 1–24, 2023

    work page 2023

  31. [31]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,”Advances in Neural Information Processing Systems, vol. 30, 2017

    work page 2017

  32. [32]

    ‘Why should i trust you?’ explaining the predictions of any classifier,

    M. T. Ribeiro, S. Singh, and C. Guestrin, “‘Why should i trust you?’ explaining the predictions of any classifier,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 1135–1144

    work page 2016

  33. [33]

    Grad-CAM: Why did you say that?

    R. R. Selvaraju, A. Das, R. Vedantam, M. Cogswell, D. Parikh, and D. Batra, “Grad-CAM: Why did you say that?”arXiv preprint arXiv:1611.07450, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  34. [34]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2921– 2929

    work page 2016

  35. [35]

    RISE: Randomized Input Sampling for Explanation of Black-box Models

    V . Petsiuk, A. Das, and K. Saenko, “Rise: Randomized input sampling for explanation of black-box models,”arXiv preprint arXiv:1806.07421, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  36. [36]

    Latent space cartography: Visual analysis of vector space embeddings,

    Y . Liu, E. Jun, Q. Li, and J. Heer, “Latent space cartography: Visual analysis of vector space embeddings,” inComputer graphics forum, vol. 38, no. 3. Wiley Online Library, 2019, pp. 67–78

    work page 2019

  37. [37]

    Closed-form factorization of latent semantics in GANs,

    Y . Shen and B. Zhou, “Closed-form factorization of latent semantics in GANs,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1532–1540

    work page 2021

  38. [38]

    L2m- gan: Learning to manipulate latent space semantics for facial attribute editing,

    G. Yang, N. Fei, M. Ding, G. Liu, Z. Lu, and T. Xiang, “L2m- gan: Learning to manipulate latent space semantics for facial attribute editing,” in2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 2950–2959

    work page 2021

  39. [39]

    Ganalyze: Toward visual definitions of cognitive image properties,

    L. Goetschalckx, A. Andonian, A. Oliva, and P. Isola, “Ganalyze: Toward visual definitions of cognitive image properties,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 5744–5753

    work page 2019

  40. [40]

    Toward a quantitative survey of dimension reduction techniques,

    M. Espadoto, R. M. Martins, A. Kerren, N. S. Hirata, and A. C. Telea, “Toward a quantitative survey of dimension reduction techniques,”IEEE Transactions on Visualization and Computer Graphics, vol. 27, no. 3, pp. 2153–2173, 2019

    work page 2019

  41. [41]

    Algorithms for hierarchical clustering: an overview, ii,

    F. Murtagh and P. Contreras, “Algorithms for hierarchical clustering: an overview, ii,”Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 7, no. 6, p. e1219, 2017

    work page 2017

  42. [42]

    Objective-based hierarchical clustering of deep embedding vectors,

    S. Naumov, G. Yaroslavtsev, and D. Avdiukhin, “Objective-based hierarchical clustering of deep embedding vectors,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 10, pp. 9055–9063, May 2021. [Online]. Available: https://ojs.aaai.org/index. php/AAAI/article/view/17094

    work page 2021

  43. [43]

    Speaker diarization using deep neural network embeddings,

    D. Garcia-Romero, D. Snyder, G. Sell, D. Povey, and A. McCree, “Speaker diarization using deep neural network embeddings,” in2017 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2017, pp. 4930–4934

    work page 2017

  44. [44]

    Supervised hierarchical clustering using graph neural networks for speaker diarization,

    P. Singh, A. Kaul, and S. Ganapathy, “Supervised hierarchical clustering using graph neural networks for speaker diarization,” inICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    work page 2023

  45. [45]

    Speaker identification and clustering using convolutional neural networks,

    Y . Lukic, C. V ogt, O. D ¨urr, and T. Stadelmann, “Speaker identification and clustering using convolutional neural networks,” in2016 IEEE 26th International Workshop on Machine Learning for Signal Processing (MLSP). IEEE, 2016, pp. 1–6

    work page 2016

  46. [46]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, “A density-based algorithm for discovering clusters in large spatial databases with noise,” ser. Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD’96). AAAI Press, 1996, p. 226–231

    work page 1996

  47. [47]

    A comprehensive survey of clustering algorithms,

    D. Xu and Y . Tian, “A comprehensive survey of clustering algorithms,” Annals of Data Science, vol. 2, pp. 165–193, 2015

    work page 2015

  48. [48]

    Comprehensive cross-hierarchy cluster agreement evaluation

    D. M. Johnson, C. Xiong, J. Gao, and J. J. Corso, “Comprehensive cross-hierarchy cluster agreement evaluation.” inAAAI (Late-Breaking Developments), 2013

    work page 2013

  49. [49]

    Hierarchical clustering algorithms for document datasets,

    Y . Zhao, G. Karypis, and U. Fayyad, “Hierarchical clustering algorithms for document datasets,”Data Mining and Knowledge Discovery, vol. 10, pp. 141–168, 2005

    work page 2005

  50. [50]

    C. D. Manning,Introduction to information retrieval. Syngress Publishing,, 2008

    work page 2008

  51. [51]

    Shortest connection networks and some generalizations,

    R. C. Prim, “Shortest connection networks and some generalizations,” The Bell System Technical Journal, vol. 36, no. 6, pp. 1389–1401, 1957

    work page 1957

  52. [52]

    On the shortest spanning subtree of a graph and the traveling salesman problem,

    J. B. Kruskal, “On the shortest spanning subtree of a graph and the traveling salesman problem,”Proceedings of the American Mathematical society, vol. 7, no. 1, pp. 48–50, 1956

    work page 1956

  53. [53]

    HDBSCAN: Hierarchical density based clustering,

    L. McInnes, J. Healy, and S. Astels, “HDBSCAN: Hierarchical density based clustering,”The Journal of Open Source Software, vol. 2, no. 11, p. 205, 2017

    work page 2017

  54. [54]

    Voxceleb2: Deep speaker recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep speaker recognition,”arXiv preprint arXiv:1806.05622, 2018

    work page arXiv 2018

  55. [55]

    Prototypical networks for few-shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few-shot learning,”Advances in Neural Information Processing Systems, vol. 30, 2017

    work page 2017

  56. [56]

    R. L. Solso, M. K. MacLin, and O. H. MacLin,Cognitive psychology. Pearson Education New Zealand, 2005

    work page 2005

  57. [57]

    Social psychology as history

    K. J. Gergen, “Social psychology as history.”Journal of personality and social psychology, vol. 26, no. 2, p. 309, 1973

    work page 1973

  58. [58]

    Paralanguage: A first approximation,

    G. L. Trager, “Paralanguage: A first approximation,”Stud. Linguist., vol. 13, pp. 1–12, 1958

    work page 1958

  59. [59]

    Speech pathology; a dynamic neurological treatment of normal speech and speech deviations

    L. E. Travis, “Speech pathology; a dynamic neurological treatment of normal speech and speech deviations.” 1931

    work page 1931

  60. [60]

    A system for describing vocal timbre in popular song

    K. Heidemann, “A system for describing vocal timbre in popular song.” Music Theory Online, vol. 22, no. 1, 2016

    work page 2016

  61. [61]

    Paralinguistic singing attribute recognition using supervised machine learning for describing the classical tenor solo singing voice in vocal pedagogy,

    Y . Xu, W. Wang, H. Cui, M. Xu, and M. Li, “Paralinguistic singing attribute recognition using supervised machine learning for describing the classical tenor solo singing voice in vocal pedagogy,”EURASIP Journal on Audio, Speech, and Music Processing, vol. 2022, no. 1, p. 8, 2022

    work page 2022