acceptodds
Under review as a conference paper at ICLR 2027

TIDE: Topological Identity Dynamics and Evidence Learning for Short-Utterance Speaker Verification

Abstract

Short-utterance speaker verification is challenging because only limited speaker information is observed. Short speech has poor phonetic coverage and misses many identity-related transitions. Its representation is therefore sparse, fragmented, and sensitive to nuisance factors. Most existing systems still compress the available frames into a fixed-dimensional embedding, without explicitly modeling missing identity structure or evidence reliability. We propose TIDE, which views a short utterance as an incomplete observation of a speaker trajectory. Frame-level features and their local temporal differences form position velocity state clouds. Multi-scale persistent homology captures stable identity structures across temporal scales. During training, a momentum long-view branch provides a richer topological target. A duration-gated residual module estimates the missing structural information from the short view without reconstructing speech. The completed topology then modulates the vector field of a neural controlled differential equation, enabling continuous identity evolution along the observed acoustic path. A thinning-consistency loss improves robustness to missing observations. For verification, each dynamic state produces directional evidence with a nonnegative weight. These contributions are accumulated over time to construct a von Mises Fisher posterior, whose mean direction represents speaker identity and whose concentration reflects the coherence of the observed evidence. Experiments on VoxCeleb under short-duration conditions demonstrate consistent improvements over strong speaker verification baselines, especially when speech duration is limited. These results show the value of jointly modeling identity structure, dynamics, and evidence for short-utterance speaker verification.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Short-Utterance Speaker Verification, Topological Representation Learning, Neural Identity Dynamics, Controlled Differential Equations, Evidential Representation Learning.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_cq6qOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TIDE for short-utterance speaker verification. It treats short speech as a partial observation of a latent speaker trajectory. The method constructs multi-scale persistent-homology representations from position–velocity feature clouds and uses a momentum long-view teacher to supervise duration-gated topology completion. The completed topology conditions the vector field of a neural controlled differential equation (CDE), while thinning consistency encourages robustness to missing observations. Dynamic states are subsequently accumulated as directional evidence to form a von Mises–Fisher (vMF) representation, with verification performed using normalized posterior overlap. Experiments on VoxCeleb1-O/E/H report improvements with ECAPA-TDNN and ERes2NetV2 encoders. Component ablations and a parameter-matched MLP control provide some support for the design. However, important questions remain regarding evaluation comparability, attribution of the gains, the behavior of the proposed scoring function, and reproducibility.

Critical strengths

Clear motivation and coherent design. Limited phonetic coverage and unreliable evidence are meaningful challenges in short-utterance verification. The proposed completion, temporal modeling, and evidence aggregation components are reasonably connected to these challenges. An interesting coupling of topology and dynamics. Conditioning the CDE vector field on completed topological representations is a potentially useful contribution. Allowing structural information to influence path integration provides a substantive architectural distinction from post-hoc feature concatenation. Explicit evidence construction. Accumulating time-weighted directional contributions gives concentration a transparent computational interpretation. The paper also appropriately limits its interpretation to task-related evidence coherence rather than claiming general epistemic uncertainty.

Critical weaknesses

The main baseline comparisons require clarification and correction. The results attributed to ERes2NetV2 in Table 1—including full-duration EERs of 0.52/0.75/1.44 and 2-second EERs of 1.43/1.60/2.96—correspond to the original paper’s ERes2NetV2 w/o (BL+BD) variant, which has 22.4M parameters. They do not correspond to the standard 17.8M model. The model identities and their relationship to the backbone used in TIDE should be stated accurately.

Matching nominal duration also does not establish matching evaluation conditions. The ERes2NetV2 reference uses full-duration enrollment and truncated test utterances, with minDCF evaluated at (P_{\mathrm{target}}=0.01). MR-RawNet uses center cropping and (P_{\mathrm{target}}=0.05). TIDE’s corresponding settings are insufficiently specified, so the relative minDCF reductions in Table 2 cannot yet be interpreted confidently.

The scoring function has an unaddressed low-evidence boundary case. Equation (26) is a valid expression for the log of normalized (L_2) posterior overlap. However, when both concentrations approach zero, both posteriors approach the uniform distribution and the score approaches its maximum value of zero, independently of their identity directions.

The experiments do not sufficiently isolate the central mechanism.

The reliability interpretation lacks direct empirical validation and reproducibility and efficiency evidence are insufficient.

Questions for authors

Why are results from ERes2NetV2 w/o (BL+BD) labeled as standard ERes2NetV2? Which architecture and checkpoint are used as TIDE’s encoder and as the reproduced baseline?

How does the system handle the case where both concentrations approach zero and Equation (26) approaches its maximum?

With identical teacher supervision and comparable capacity, how does persistent topology compare with non-topological summaries or ordinary long–short feature distillation? What are the separate contributions of (H_0) and (H_1)?

Does the neural CDE outperform discrete temporal models supplied with the same topology? Do temporal-order perturbations support the proposed trajectory interpretation?

2 comments
AuthorsOpenReview ↗

Response on Evaluation Protocols, Score Boundaries, and Mechanism Validation

Thank you for recognizing the joint design. We added the following corrections and validations. EERs below are O/E/H macro-averages.

1. Encoder identities and comparable protocols (W1-W2/Q1).

We corrected the published rows to ERes2NetV2 w/o (BL+BD), 22.4M parameters (Tables 1-2/D1). Both the reproduced baseline and TIDE encoder use standard 17.8M ERes2NetV2; TIDE totals 20.16M. New matched runs train from scratch with shared encoder initialization per seed (2027/2028/2029) and development-selected checkpoints, using VoxCeleb2-dev with a speaker-disjoint 95/5 split, 80-dimensional Fbank, shared augmentation, and 80,000 Adam updates.

Published comparisons are explicitly external context. New Table D2 uses identical saved crops and official trial identities, with minDCF priors 0.01 and 0.05 reported separately. At 16 kHz, target DD uses L=16000DL=16000D samples, start max⁡(0,⌊(N−L)/2⌋)\max(0,\lfloor(N-L)/2\rfloor) and length min⁡(N,L)\min(N,L) for recording length NN; shorter inputs retain actual speech, with VAD disabled. SS-D crops both sides; FS-D uses full enrollment and short test.

At SS-2, baseline/TIDE EERs are 2.20%/1.76% for ECAPA and 2.05%/1.62% for ERes2NetV2. Their minDCF multiplied by 100 improves from 21.30 to 16.98 and 19.98 to 15.62 at prior 0.01, and from 13.75 to 10.90 and 12.90 to 10.05 at prior 0.05. These comparisons hold encoder and evaluation conditions fixed.

2. Low-evidence boundary and explicit handling (W3/Q2).

Original Eq. (26), now Eq. (27), measures normalized L2L_2 density overlap: two uniform posteriors attain its maximum zero. We now explicitly distinguish posterior-shape agreement from evidence for a shared identity and add the uniform-reference score (Eqs. F5-F7): sU=log⁡Cd(κe)+log⁡Cd(κt)−log⁡Cd(∥ηe+ηt∥2)−log⁡Cd(0).s_U=\log C_d(\kappa_e)+\log C_d(\kappa_t)-\log C_d(\|\eta_e+\eta_t\|_2)-\log C_d(0). Here η\eta is the vMF natural parameter, κ=∥η∥2\kappa=\|\eta\|_2, and CdC_d is the normalizer. Exactly sU(0,η)=sU(η,0)=0s_U(0,\eta)=s_U(\eta,0)=0; aligned concentrated posteriors can score above zero. Development-fitted positive-scale calibration p=σ(asU+logit⁡π)p=\sigma(a s_U+\operatorname{logit}\pi), a>0a>0, maps uniform input to prior π\pi. With unit costs and π=0.01\pi=0.01, prior-only probability 0.01 is below the 0.5 acceptance threshold.

Original main-table results retain NML. Added Table F5 separately reports: original NML, 4.18%/1.76% EER at 1/2 s; same-checkpoint uniform-reference rescoring, 4.15%/1.75%; uniform-reference pair-objective retraining, 4.06%/1.70%. Low-evidence FAR decreases from 6.8% to 4.6% and 3.9%, respectively. All nonempty trials remain in evaluation; the subset is development-defined. The 768-case check agrees with an 80-digit reference within 2.753×10−122.753\times10^{-12} (Table F1).

3. Topology versus multiscale capacity and teacher supervision (W4/Q3).

New Table E1 controls encoder, normalized clouds, three 128-dimensional tokens, completion, corresponding EMA long-view targets, CDE, evidence and update budget. SS-1/SS-2 EERs are: statistics 4.63%/2.03% (16.98M); DeepSets 4.49%/1.96% (17.16M); pair-distance summaries 4.42%/1.92% (17.04M); persistent topology 4.18%/1.76% (17.09M). Ordinary EMA acoustic-feature distillation gives 4.56%/2.00%; topology without the long-view target gives 4.78%/2.05% while retaining completion capacity.

Table E2 holds parameter count at 17.09M: H0H_0-only/H1H_1-only/joint EERs are 4.42%/4.57%/4.18% at 1 s and 1.88%/1.96%/1.76% at 2 s. Both homology orders contribute under the same downstream width and training budget.

4. CDE and temporal-order evidence (W4/Q4).

With the same completed topology, Table E3 gives GRU/TCN/CDE EERs of 4.43%/4.36%/4.18% at 1 s and 1.90%/1.85%/1.76% at 2 s (17.06/17.12/17.09M parameters).

In added Table E4, permutations occur after frame encoding on an increasing time grid. Keeping topology fixed, interior permutation raises CDE EER from 4.18% to 5.08%, while static pooling remains 4.95%. The same permutation with velocity/topology recomputed gives 5.55%. With 25%/50% internal-knot removal, CDE EERs are 4.39%/4.89%, versus GRU 4.91%/5.70%. These interventions support order-sensitive integration and robustness to missing observations.

5. Reliability, reproducibility and efficiency (W5).

New Table F3 adds log⁡(1+κe)+log⁡(1+κt)\log(1+\kappa_e)+\log(1+\kappa_t) to the same duration-plus-score error predictor: AUROC improves from 0.733 to 0.771 at clean SS-2, 0.702 to 0.753 at 10 dB noise, and 0.711 to 0.759 with held-out RIR. Predictors and operating thresholds use development data only.

Appendix C now specifies scales 1/2/4, 32 landmarks, EMA 0.999, 64/192 control/state dimensions, two RK4 substeps, and loss weights λtop=1\lambda_{\rm top}=1, λdyn=λev=λpair=0.1\lambda_{\rm dyn}=\lambda_{\rm ev}=\lambda_{\rm pair}=0.1. Table G2 adds complete pipeline timing: on A100, FP32, batch 1, ECAPA baseline/TIDE at 2 s use 94.7/124.0 ms and 1500/1770 MB; TIDE includes 18.6 ms CPU topology time. ERes2NetV2 baseline/TIDE use 200/204 ms.

Based on the rebuttal provided by the authors, I believe my main concern of this paper has been solved. I change the score to weak acceptance.

Reviewer_YDXnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

TIDE targets speaker verification with short utterances, where sparse phonetic coverage makes identity evidence less reliable. It combines topological descriptors of frame-level structure with a continuous-time evidence model and a directional classification objective. The reported experiments and ablations show encouraging short-duration verification performance, including degradation when the topological representation is moved outside the evidence model. The main question for me is whether the gains are specifically due to the proposed topology and whether the experimental setup is sufficiently documented to reproduce that conclusion.

Critical strengths
  1. The short-utterance setting is important, and the paper offers a coherent reason for studying how identity evidence develops over time rather than relying only on a final pooled embedding.
  2. The ablation evidence is useful. Table 4 (p. 9) reports a meaningful increase in average EER when topology is moved outside the CDE, which supports the relevance of their joint design.
Critical weaknesses
  1. The experiments do not cleanly isolate the value of persistent topology. A matched, non-topological descriptor with comparable dimensionality and computational budget would help show that the gain comes from the claimed topological information rather than from an additional multiscale feature stream or model capacity. The parameter-matched MLP study in the appendix is helpful, but it does not fully resolve this comparison.
  2. Several implementation and protocol choices needed for reproduction are unclear, including the scales and landmarks, momentum update, loss weights, CDE settings, and short-utterance crop and trial construction. This limits confidence in the reported comparison, especially when cited baseline systems differ in training data or front ends. The presentation also needs repair: the method figures are dense (appear to be AI generated), and Section 3.3 (p. 8) contains unresolved “Figure ??” references.
Questions for authors
  1. Can the authors provide a controlled non-topological descriptor baseline with the same front end, feature dimension, training setup, and parameter budget as TIDE?
  2. Can the authors release or specify the missing hyperparameters and exact 1 s/2 s crop and trial protocol, including how each reported baseline was trained and evaluated? These details would materially affect reproducibility and the strength of the comparison.
2 comments
AuthorsOpenReview ↗

Response on Topology Controls, Reproducibility, and Presentation

Thank you for recognizing the joint design. We added the following controls, specifications, and presentation fixes.

1. Isolating persistent topology (W1/Q1).

New Table E1 replaces only the descriptor stream while retaining the same encoder, normalized position-velocity clouds, three 128-dimensional tokens, duration-gated completion, corresponding EMA long-view targets, CDE, evidence head, and 80,000-update budget. Controls are statistics (mean/std/local differences), DeepSets (point MLP, mean/max pooling), and distance summaries (32 quantiles plus 32 neighborhood counts).

SS-1/SS-2 macro-average EERs over VoxCeleb1-O/E/H are: statistics, 4.63%/2.03% (16.98M parameters); DeepSets, 4.49%/1.96% (17.16M); distance summaries, 4.42%/1.92% (17.04M); persistent topology, 4.18%/1.76% (17.09M). The 2 s latencies are 111 ms for DeepSets and 124 ms for TIDE (A100, FP32, batch 1).

Ordinary feature distillation with the same EMA long-view signal gives 4.56%/2.00%; topology without the long-view target gives 4.78%/2.05% while retaining completion capacity (Table E1). Thus the comparison controls both an additional multiscale stream and privileged long-view supervision.

2. Exact implementation and evaluation (W2/Q2).

Added in Tables C1-C3 and Appendix D:

  • Structure: pooling supports/strides are 1, 2, 4 frames, with at most 32 genuine landmarks per scale. For an nn-point cloud, L=min⁡(n,32)>1L=\min(n,32)>1 uses indices ⌊j(n−1)/(L−1)+1/2⌋\lfloor j(n-1)/(L-1)+1/2\rfloor, j=0,…,L−1j=0,\ldots,L-1. Endpoints are included; shorter clouds retain all points without duplication. H0H_0 and H1H_1 use Vietoris-Rips filtration over F2\mathbb{F}_2, including dimension-two simplices, and 64 Gaussian coordinates per order, projected to 128 dimensions per scale (initial Gaussian width 0.25). State projection is 64-dimensional and velocity weight is 0.5.
  • Momentum/losses: initialize the teacher from the student and update after every optimizer step as ϑˉ←0.999ϑˉ+0.001ϑ\bar\vartheta\leftarrow0.999\bar\vartheta+0.001\vartheta. EMA includes the encoder, state projection, persistence vectorization, token projection and target contrastive projector; targets are stopped, normalization buffers copied, and teacher dropout disabled. We use λtop=1\lambda_{\rm top}=1, λtcon=0.1\lambda_{\rm tcon}=0.1, λdyn=λev=λpair=0.1\lambda_{\rm dyn}=\lambda_{\rm ev}=\lambda_{\rm pair}=0.1, βKL=10−4\beta_{\rm KL}=10^{-4}, contrastive temperature 0.1, AAM margin/scale 0.2/30, and evidence rate bound 1024/s.
  • CDE/completion: control/state dimensions are 64/192; the vector field has two width-128 tanh layers, with four-head topology attention at every RK4 evaluation. Piecewise-linear controls use two RK4 substeps per interval and direct backpropagation. Initialization uses the first control value. Completion uses two width-128, four-head attention blocks; Dref=5D_{\rm ref}=5 s. Internal-knot thinning probability is 0.25, preserving endpoints/timestamps, with eight trajectory anchors.
  • Training: all new matched runs use VoxCeleb2-dev, a speaker-disjoint 95/5 train/development split, 16 kHz mono audio, 80-dimensional Fbank, 25/10 ms window/hop, and FFT 512. Adam (betas 0.9/0.999) uses peak LR 5×10−45\times10^{-4}, 5,000 warm-up updates, cosine decay to 5×10−65\times10^{-6}, weight decay 10−410^{-4}, and gradient clipping 1. Batches contain 16 speakers x 4 distinct utterances; seeds are 2027/2028/2029. Noise probability is 0.5 with SNR 0-20 dB; RIR probability is 0.5. Each variant uses 80,000 updates with shared encoder initialization per seed.
  • Crops/trials: training uses a genuine long view up to 8 s and valid nested 1/2/3/5 s short views, with coupled augmentation. Primary evaluation disables VAD. For NN recording samples and target DD seconds, L=16000DL=16000D samples, crop start is max⁡(0,⌊(N−L)/2⌋)\max(0,\lfloor(N-L)/2\rfloor) and length is min⁡(N,L)\min(N,L). Thus 1/2 s request 16,000/32,000 samples; shorter recordings use actual duration without repetition. Saved offsets and official VoxCeleb1-O/E/H trial identities are shared across systems. SS-D crops both sides; FS-D keeps full enrollment; FF keeps both full. Training verification pairs are balanced, with distinct recordings for same-speaker positives. Selection/calibration use disjoint development speakers; minDCF priors 0.01/0.05 are reported separately.

Published results are identified as external references, including the 22.4M ERes2NetV2 w/o (BL+BD), separate from the 17.8M matched encoder. New Table D2 gives SS-1/SS-2 O/E/H macro-average EER changes of 4.95/2.20% to 4.18/1.76% for ECAPA and 4.76/2.05% to 3.85/1.62% for ERes2NetV2 under the common protocol.

3. Presentation (W2).

Section 3.3 now refers explicitly to Figure 2(d) and Figure 3. Added module interfaces and implementation rules explain Figure 1's flow: observed state clouds, persistent tokens, supervised completion, topology-conditioned integration, and physical-time vMF evidence. Algorithm 1 provides the corresponding training/inference sequence.

AuthorsOpenReview ↗

Follow-up on Added Controls and Reproducibility Details

Dear Reviewer YDXn,

Thank you again for your time and constructive review. We would like to follow up on our point-by-point response and revised manuscript.

We have added the non-topological descriptor controls you requested, using the same front end, token dimensions, EMA supervision, and training budget. At 2 s, DeepSets gives 1.96% EER with 17.16M parameters, compared with 1.76% for TIDE with 17.09M parameters, averaged over VoxCeleb1-O/E/H. We have also specified the missing hyperparameters and exact 1 s/2 s crop and trial protocols, and corrected the figure references.

We would appreciate knowing whether these additions address your concerns. If any questions remain, we would be happy to provide further clarification during the discussion period.

Thank you for your time and consideration.

Reviewer_AB6jOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TIDE, a framework for short-utterance speaker verification that models limited speech as incomplete observations of speaker identity trajectories. It combines multi-scale persistent homology and duration-gated topology completion with topology-conditioned neural controlled differential equations to capture continuous identity dynamics. A von Mises–Fisher posterior further integrates directional speaker evidence and its reliability. Experiments on VoxCeleb1-O/E/H demonstrate competitive verification performance across different utterance durations, particularly for short speech. Evaluations with ECAPA-TDNN and ERes2NetV2 encoders, together with component ablations, support the effectiveness of the proposed framework.

Critical strengths
  1. TIDE integrates multi-scale persistent homology with neural controlled differential equations (CDEs), allowing topological information to directly modulate identity dynamics rather than serving as a static feature descriptor. The duration-gated topology completion mechanism further leverages long-utterance supervision to compensate for missing structural information in short speech. This is a significant strength because it provides a technically distinctive approach to modeling incomplete temporal observations and their evolution.

  2. TIDE introduces a von Mises–Fisher (vMF) posterior that jointly represents speaker identity direction and evidence concentration. By accumulating directional evidence in physical time, the framework distinguishes the amount of available speech from the coherence of identity evidence. This is consequential because it provides a mathematically motivated mechanism for incorporating evidence reliability directly into speaker verification, rather than relying solely on deterministic embeddings and cosine similarity.

Critical weaknesses

TIDE assumes that persistent topological features extracted from position–velocity state clouds capture speaker-specific identity structures. However, these features may also reflect phonetic content, speaking style, or recording conditions. Although the ablation studies demonstrate the effectiveness of topology-based modeling, they do not directly establish that the extracted structures are primarily associated with speaker identity rather than nuisance factors.

  1. The experimental comparisons rely partly on previously reported results from methods with different acoustic front ends, training configurations, and evaluation protocols. Although TIDE achieves substantial improvements in several settings, its advantages are not consistent across all matched-duration comparisons. For example, at two seconds on VoxCeleb1-E, TIDE achieves an EER of 1.66%, compared with 1.60% for ERes2NetV2. Without more tightly controlled comparisons under identical training and evaluation conditions, it remains difficult to determine how much of the observed improvement can be attributed specifically to the proposed topology–dynamics–evidence framework.
Questions for authors
  1. What evidence supports the claim that the learned topology primarily represents speaker identity rather than these nuisance factors? Could the authors clarify whether its identity-discriminative properties remain consistent across different phonetic contents or recording conditions?

2.Could the authors clarify how differences in training data, augmentation, feature extraction, and evaluation protocols affect these comparisons? In particular, can the performance gains be demonstrated under matched encoder and training settings against strong speaker verification baselines?

  1. Could the authors clarify whether the predicted concentration correlates with actual verification accuracy across different utterance durations and acoustic conditions?
2 comments
AuthorsOpenReview ↗

Response on Speaker Discriminability, Matched Comparisons, and Evidence Reliability

We thank the reviewer for recognizing the topology-conditioned dynamics and physical-time evidence formulation. We have added experiments addressing all three concerns.

1. Speaker identity versus nuisance factors (W1/Q1).

Speaker supervision explicitly shapes topology: Eq. (14) contrasts completed short-view tokens with long-view targets using same-speaker positives P(i)={p:yp=yi}\mathcal{P}(i)=\{p:y_p=y_i\}.

We added frozen probes on 100 speakers, using source-video-disjoint splits of 40/10/20 utterances per speaker for probe training/development/evaluation (2,000 evaluation samples). Representations use 128-dimensional PCA determined from probe-training data and the same linear probe. Speaker balanced accuracy is 65.8% for DeepSets, 70.9% for observed topology, 75.6% for completed topology, and 80.3% for the final evidence direction. From DeepSets to completed topology, ASR-content accuracy decreases from 47.6% to 40.5%, and acoustic-condition accuracy from 68.4% to 57.9% (Table E6). Completion therefore increases speaker discriminability while reducing nuisance decodability.

We also added content/condition verification (Table E5). SS-2 macro-average EERs over VoxCeleb1-O/E/H, listed as DeepSets versus TIDE, are: within-content/video-disjoint, 1.93% versus 1.74%; cross-content/video-disjoint, 2.54% versus 2.14%; content-matched, 2.21% versus 1.95%; 10 dB noise, 3.71% versus 3.24%; held-out RIR, 3.25% versus 2.88%. Content-matched evaluation constrains impostor content while retaining genuine trials. These results support speaker-selective structure across content and recording changes.

2. Matched comparisons and framework attribution (W2/Q2).

The cited 1.66% and 1.60% compare ECAPA-TIDE with the published 22.4M ERes2NetV2 w/o (BL+BD), involving different encoders and training records. We now identify the published variants and add within-encoder comparisons (Tables C2, D1-D2).

New matched runs share VoxCeleb2-dev with a speaker-disjoint 95/5 train/development split, 16 kHz audio, 80-dimensional Fbank features (25 ms window/10 ms hop), encoder initialization per seed, Adam, and 80,000 updates. They share noise augmentation (probability 0.5, SNR 0-20 dB), RIR augmentation (probability 0.5), saved evaluation crops and official trial identities. Checkpoint selection and calibration use development data only.

For SS-D, both sides are cropped to D seconds. At SS-1/SS-2/SS-5, ECAPA pooling versus ECAPA-TIDE EERs are 4.95/4.18%, 2.20/1.76%, and 1.73/1.38%; standard ERes2NetV2 versus ERes2NetV2-TIDE gives 4.76/3.85%, 2.05/1.62%, and 1.61/1.27% (O/E/H macro-averages). At SS-2, relative EER reductions are 20.0% and 21.0%. MinDCF multiplied by 100 at prior 0.01 improves from 21.30 to 16.98 and from 19.98 to 15.62, respectively.

New descriptor controls retain completion, EMA supervision, CDE, evidence, and the update budget: DeepSets gives 4.49%/1.96% at SS-1/SS-2, ordinary feature distillation 4.56%/2.00%, and persistent topology 4.18%/1.76% (Table E1). With identical topology input, a TCN gives 4.36%/1.85% versus the CDE's 4.18%/1.76% (Table E3). The parameter-matched MLP and ECAPA-TIDE both have 17.09M parameters but give 3.92% versus 3.41% on 1 s VoxCeleb1-O (Table 5). These controls separate framework gains from encoder choice, teacher supervision, and parameter growth.

3. Concentration and verification reliability (Q3).

Yes. We added concentration/error and calibrated-reliability analyses for the reported NML system (Tables F2-F4). Eqs. (24), (26), and (F2) give η=∑iδiaiui,κ=∥η∥2,E=∑iδiai,κ≤E.\eta=\sum_i\delta_i a_i u_i,\quad \kappa=\|\eta\|_2,\quad E=\sum_i\delta_i a_i,\quad \kappa\le E. Here δi\delta_i weights physical time, aia_i is the evidence rate, and uiu_i is a unit direction. Coherent directions reinforce concentration; conflicting directions reduce it.

Pair quality is Qe,t=Ad(κe)Ad(κt)Q_{e,t}=A_d(\kappa_e)A_d(\kappa_t), where Ad(κ)=Id/2(κ)/Id/2−1(κ)A_d(\kappa)=I_{d/2}(\kappa)/I_{d/2-1}(\kappa) and d=192d=192. At clean SS-1/SS-2/SS-5/full-full, median concentration is 74/156/402/765, with EER 4.18/1.76/1.38/0.49%. At SS-2, 10 dB noise and held-out RIR reduce median concentration to 93 and 110, with EER 3.24% and 2.88%.

Within SS-2 clean/noise/RIR, quality/error Spearman correlations are -0.12/-0.17/-0.15. Adding log⁡(1+κe)+log⁡(1+κt)\log(1+\kappa_e)+\log(1+\kappa_t) to the same duration-plus-score error predictor improves error AUROC from 0.733 to 0.771, 0.702 to 0.753, and 0.711 to 0.759, respectively. Thus concentration contributes reliability information beyond duration and score magnitude. Predictors, calibration and operating thresholds use development data only.

With equal affine-calibration capacity, TIDE cosine versus NML gives SS-2 EER 1.87% versus 1.76%, NLL 0.0125 versus 0.0109, ECE 0.0018 versus 0.0013, and risk-coverage AURC 0.00076 versus 0.00061 (Table F4).

Thank you for the additional experiments. The speaker-discriminability probes, matched within-encoder comparisons, and concentration-based reliability analyses substantially address my concerns. Although the probes do not establish complete separation from nuisance factors, the evidence strengthens the paper’s technical and empirical contributions. I maintain my rating of 3

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.