PaperlayerView exact revision on arXiv

Source-linked AI summary

VoiceNet: Fine-Grained Voice Understanding Beyond Emotion at Scale

Christoph Schuhmann, Robert Kaczmarczyk, Gollam Rabby, Felix Friedrich, Maurice Kraus, Gijs Wijngaard, Kourosh Nadi, Huu Nguyen, Kristian Kersting, Sören Auer

arXiv:2609.32016v1cs.SDcs.AIeess.AS

TL;DR

Public emotion benchmarks cover few basic categories, and no public benchmark scores talking-style attributes alongside emotion. VoiceNet combines fine-grained emotion and talking-style evaluation with emotion-annotated speech resources; VoiceCLAP-Large outperforms existing CLAP baselines on VoiceNet-Emo.

  • Problem

    Public emotion benchmarks use few basic categories, while no public benchmark scores talking-style attributes alongside emotion on in-the-wild speech.

    Source: §1 · pp. 1–2

  • Method

    The paper introduces VoiceNet-Emo with 40 emotions and three expert raters per item, preliminary VoiceNet-Ext with 57 talking-style attributes, and Emolia resources enriched with MOSS-Audio annotations.

    Source: §1 · p. 2, §3.3 · p. 5

  • Results

    VoiceCLAP-Large reaches 0.702 VoiceNet-Emo per-prompt balanced accuracy and mean Spearman ρ=0.372, surpassing speech-specialised CLSP by over +0.31 ρ.

    Source: §5 · p. 8, §6 · p. 10

  • Takeaways & Limitations

    General-audio captions and speech-aware pretraining alone do not provide dense vocal-style supervision, while contrastive finetuning on Emolia-Balanced adds +0.039 bal@pp to Omni-Embedding bases.

    Source: §5 · p. 9

    VoiceNet-Ext is preliminary and uses binary confirmations of level matches rather than full ordinal ratings; source corpora have non-uniform geographic, demographic, and recording-environment coverage.

    Source: §3.3 · p. 5, §5 · p. 10

Abstract

from arXiv · show

Expressive speech synthesis has outpaced expressive speech perception: systems now render fine-grained vocal performances that no public benchmark can score. Most benchmarks for this inverse problem stop at six to nine basic emotion categories, largely on acted speech. This paper introduces VoiceNet, a human-annotated representation-level benchmark for voice performance understanding on permissively-licensed in-the-wild speech. VoiceNet has two subsets: VoiceNet-Emo applies a 40-emotion taxonomy with three expert ratings per item, and VoiceNet-Ext, a preliminary subset, scores 57 talking-style attributes including speaking rate, vocal tension, breathiness, and register. The paper also releases Emolia, an emotion-annotated version of the Emilia corpus, with a curated rebalanced subset enriched by dense MOSS-Audio Thinking annotations. Two voice-text contrastive baselines train on this data: a 110M-parameter VoiceCLAP-Small for fast large-scale data filtering and a 7B VoiceCLAP-Large for state-of-the-art performance. Both outperform existing CLAP baselines, which sit near chance on VoiceNet-Emo. On VoiceNet-Emo, VoiceCLAP-Large aligns more closely with the aggregate expert consensus than individual experts agree with one another: a comparison against the majority label rather than evidence of surpassing human emotion perception. All systems evaluated here are voice-text embedding models: VoiceNet scores representation-level attribute recognition and retrieval, not end-to-end spoken-dialogue behaviour. Clustering and filtering uncurated speech corpora into subsets that span diverse talking styles and emotions remains an open challenge; VoiceCLAP embeddings offer a promising tool for this task. VoiceNet, Emolia, and VoiceCLAP are publicly available for research use.

1 Introduction

Public speech benchmarks offer limited coverage of fine-grained emotion and lack talking-style evaluation on in-the-wild speech. VoiceNet addresses these gaps with a human-annotated benchmark, associated training resources, and voice-text models.

Motivation

  • Public emotion benchmarks cover only 6 to 9 basic emotions, mostly on acted speech, while no public benchmark evaluates talking-style attributes alongside emotion.The missing attributes include speaking rate, vocal tension, breathiness, register, chunking, timbre, and recording context.Source: §1 · pp. 1–2
  • The 40-emotion taxonomy is validated on synthetic audio and applies to permissively licensed real-human speech; MOSS-Audio can produce structured voice annotations at scale.Existing CLAP and voice-text models were trained on general-audio captions rather than dense vocal-style descriptions.Source: §1 · p. 2

Contributions

  • VoiceNet pairs a 40-emotion benchmark with a 57-attribute talking-style subset on permissively licensed, in-the-wild speech.The paper also releases Emolia resources and VoiceCLAP models; existing general-audio CLAPs are near chance on VoiceNet-Emo.Source: §1 · p. 2

Scope

  • VoiceNet evaluates voice-text embedding models’ ranking of fine-grained descriptions against speech, not end-to-end assistants or conversational behavior.Generative audio-language models are also out of scope because they require a different scoring interface than cosine similarity.Source: §1 · p. 2

Open challenge

  • Clustering and filtering uncurated speech into subsets spanning diverse styles and emotions remains open; VoiceCLAP embeddings offer a practical handle through nearest-neighbor search.The paper demonstrates this use case with an interactive browser of Emolia-Balanced.Source: §1 · p. 2

2 Related Work

Prior speech-emotion benchmarks commonly use a small set of acted-speech labels, while vocal-expression research and affective science describe broader and more graded spaces. VoiceNet extends this landscape to fine-grained emotions, talking-style attributes, and in-the-wild speech.

Speech emotion benchmarks

  • Public SER datasets generally use six to nine acted emotion classes, while vocal-expression studies map a much larger space of distinct emotions.Aggregations broaden dataset and language coverage but inherit acted speech and narrow taxonomies; MSP-PODCAST is one of few large in-the-wild corpora.Source: §2 · pp. 2–3

Annotation taxonomies and dimensional models

  • VoiceNet ports the 40-emotion taxonomy from synthetic EMONET-VOICE to in-the-wild human speech and adds 57 talking-style attributes scored with contrastive cosine similarity.VoiceNet-Ext uses binary level rubrics as a complementary form of supervision beyond emotion categories.Source: §2 · p. 3

Voice-text contrastive models

  • General-audio CLAPs use sound-event captions rather than dense vocal-style annotations, and the paper reports that they sit at chance on VoiceNet-Emo.Single-tower Omni-Embedding models reach approximately 0.65 on VoiceNet-Emo zero-shot.Source: §2 · p. 3

Annotation taxonomies and dimensional models

  • Affective science treats emotion as context-dependent and graded, yet most benchmarks assign one discrete label per clip.Prior intensity-annotation work reports low crowdsourced agreement.Source: §2 · p. 3

3 VOICENET Suite

The VoiceNet suite combines taxonomies, annotated training resources, a human-rated in-the-wild benchmark, and two voice-text contrastive models. Its benchmark subsets differ in maturity, with VoiceNet-Ext explicitly preliminary.

3 VOICENET Suite

  • The suite comprises a 40-emotion and 57-attribute taxonomy, annotated training corpora, two benchmark subsets at different maturity levels, and two VoiceCLAP models.VoiceNet-Emo is near-complete and expert-rated; VoiceNet-Ext is preliminary, with inter-rater agreement at chance.Source: §3 · p. 3

3.1 Taxonomies

  • The suite’s taxonomy spans 40 emotions and 57 talking-style attributes grouped across speaker traits, affect, delivery, vocal quality, resonance, recording context, and style.The talking-style set excludes vocal bursts, fine-grained emotion, accent, and language from the 61-value MOSS-Audio schema for specified redundancy or category reasons.Source: §3.1 · p. 3

3.2 EMOLIA, EMOLIA-BALANCED, and the MOSS-AUDIO Annotation Pipeline

  • 71.78M Emilia clips are annotated in Emolia with 40 emotion scores, captions, and speaker-timbre embeddings; Emolia-Balanced is a curated 5.26M-clip subset.Emolia-Balanced is rebalanced across 40 emotions and 3,000 speaker clusters; the MOSS-Audio annotation suite totals 8.64M clips across four corpora.Source: §3.2 · p. 4
  • MOSS-Audio Thinking processes 18 prompt groups per clip to produce 61 attribute values for contrastive training; retained reasoning traces are not used in that training.The training mixture also includes smaller voice corpora that broaden the contrastive negative distribution.Source: §3.2 · p. 4

3.3 VOICENET: A Human-Annotated In-the-Wild Benchmark

  • 7,988 VoiceNet-Emo audio-emotion pairs use a 40-emotion taxonomy and mostly receive three independent expert ratings on a three-point presence scale.Five task types balance emotion presence; the classifier used for preselection is a sampling prior, while released labels come from independent expert majority votes.Source: §3.3 · pp. 4–5
  • 18.5k VoiceNet-Ext pairs cover 57 attributes, but three-rater agreement is near chance and no attribute meets the prespecified reliable-core rule.The paper therefore treats all VoiceNet-Ext scores as preliminary and the attributes as exploratory perceptual probes.Source: §3.3 · p. 5

3.4 VOICECLAP: Voice-Text Contrastive Models

  • VoiceCLAP-Small is a 110M-parameter dual-tower model trained with SigLIP, while VoiceCLAP-Large is a rank-16 LoRA fine-tune of a 7B Omni-Embedding model trained with InfoNCE.The headline table reports results on items with at least two human raters, using a multi-rater majority rather than a single annotator’s call.Source: §3.4 · p. 5, Table 1 · p. 6

4 Experiments

The experiments compare two VoiceCLAP finetunes with zero-shot Omni-Embedding bases and general-audio CLAPs using VoiceNet and external speech-emotion benchmarks. VoiceCLAP-Large leads on both VoiceNet subsets and wins two of four external benchmarks, while VoiceNet-Ext results remain exploratory given weak human consistency.

4 Experiments

  • The harness L2-normalizes audio and prompt embeddings, then reports thresholded balanced accuracy and threshold-free per-prompt mean Spearman ρ as its primary unbiased ranking statistic.Per-prompt threshold scores are oracle-calibrated because thresholds are fitted and scored on the same labels; five-fold clip-grouped held-out calibration is also reported.Source: §4 · p. 6

4.1 Where the Speech-Emotion Gap Lives

  • 0.702 bal@pp puts VoiceCLAP-Large atop VoiceNet-Emo, +0.039 over its zero-shot base; it tops every column on both subsets, with bootstrap support for top-versus-runner-up gaps.VoiceCLAP-Small, at 110M parameters, lands within ∼0.03 of its 7B sibling and clears every general-audio CLAP.Source: §4.1 · p. 7
  • VoiceCLAP-Large’s 0.7021 bal@pp exceeds the 0.572 leave-one-out expert reference on VoiceNet-Emo, indicating closer alignment with aggregate consensus, not surpassing human emotion perception.Held-out bal@pp is 0.6820 for VoiceCLAP-Large and 0.6412 for its LCO-Omni-7B base, a paired lead of +0.0409.Source: §4.1 · p. 7

4.2 Per-Emotion Performance on VOICENET-EMO

  • At least one model exceeds ρ=0.5 on distress, anger, impatience, and sadness, while VoiceCLAP-Large wins or ranks second on most emotions but trails LCO bases on three categories.Doubt, Sourness, Interest, Sexual Lust, and Jealousy/Envy have best ρ ≤0.29 across all eleven models.Source: §4.2 · p. 7

4.3 Per-Attribute Performance on VOICENET-EXT

  • VoiceNet-Ext is harder than VoiceNet-Emo: mean per-prompt ρ is 0.15 versus 0.37 for VoiceCLAP-Large and 0.11 versus 0.32 for VoiceCLAP-Small.It has about 46 pairs per prompt versus about 200 per emotion, and its human consistency reference is at chance; no attribute reaches κ ≥0.20, so aggregate results are exploratory.Source: §4.3 · pp. 7–8
  • None of four paired high-minus-low agreement interactions for VoiceCLAP-Large’s advantage over LCO-Embedding-Omni-7B excludes zero, so the data do not show its edge concentrated in higher-agreement categories.The result applies to both emotions and attributes.Source: §4.3 · p. 8

4.4 Cross-Dataset Evaluation on Public Speech-Emotion Benchmarks

  • VoiceCLAP-Large wins two of four external benchmarks: it reaches 0.321 on IEMOCAP and 0.511 on CREMA-D, while ranking second on EmoNet-Voice and RAVDESS.On EmoNet-Voice, general-audio CLAPs score 0.024–0.036 (chance is 0.025); VoiceCLAP-Large scores 0.155, behind LCO-Embedding-Omni-7B at 0.167.Source: §4.4 · p. 8
  • Follow-up analysis attributes VoiceCLAP-Large’s EmoNet-Voice and RAVDESS deficits against its base to class-specific boundary shifts, while RAVDESS macro F1 rises from 0.285 to 0.330 despite lower micro accuracy.Ten of eleven models were evaluated across the four datasets using cosine-similarity argmax predictions and top-1 accuracy.Source: §4.4 · p. 8

5 Discussion

VoiceCLAP models outperform general-audio CLAP baselines on VoiceNet-Emo, while talking-style attribution remains harder and less reliable. Results are bounded by subjective emotion labels, preliminary VoiceNet-Ext annotations, and representation-level evaluation.

5 Discussion

  • All seven CLAP baselines stay at |ρ| < 0.1 on VoiceNet-Emo; VoiceCLAP-Large exceeds CLSP by +0.311 ρ and +0.135 bal@pp.VoiceCLAP-Small exceeds CLSP by +0.257 and +0.108, respectively.Source: §5 · pp. 8–9
  • VOICECLAP-LARGE reaches ρ=0.372, above the 75th-percentile individual-differences correlation of r=.29, but these values reflect subjective rank-ordered ratings.Per-emotion peaks are 0.57–0.63 for distress, impatience, anger, and sadness.Source: §5 · p. 9
  • VOICECLAP-LARGE averages ρ=0.15 per VOICENET-EXT prompt versus 0.37 on VOICENET-EMO; EXT agreement is at chance across 5,583 three-rater items.The authors identify rubric definitions and anchor-example rater training as next steps, and characterize EXT as preliminary.Source: §5 · p. 9
  • The evaluation covers voice-text embedding models only, and MOSS-Audio annotations are model-generated with human-audit evidence that bounds rather than establishes label quality.VOICENET-EXT agreement remains at chance, none of its attributes passes the reliable-core rule, and source corpora have non-uniform geographic, demographic, and recording-environment coverage.Source: §5 · pp. 9–10

6 Conclusion

The paper releases a human-annotated benchmark for fine-grained emotion and talking-style understanding, alongside annotated training corpora and voice-text contrastive models. VoiceCLAP outperforms general-audio CLAP baselines on VoiceNet-Emo, while corpus filtering across diverse styles remains open.

6 Conclusion

  • VOICENET is a benchmark on permissively-licensed real human speech combining a 40-emotion expert-rated subset (VOICENET-EMO) with a 57-attribute talking-style subset (VOICENET-EXT).Source: §6 · p. 10
  • The released resources include Emolia, the curated Emolia-Balanced subset with dense MOSS-Audio annotations, three sibling annotated corpora, and 110M and 7B VoiceCLAP models..Source: §6 · p. 10
  • VoiceCLAP-Small and VoiceCLAP-Large reach 0.675 and 0.702 VoiceNet-Emo per-prompt balanced accuracy, respectively, while all seven evaluated CLAP baselines remain near chance.Clustering and filtering uncurated speech into subsets spanning diverse talking styles and emotions remains an open challenge; VoiceCLAP embeddings are presented as a promising tool.Source: §6 · p. 10

AI use statement

The paper uses AI models for annotation and research support, while released labels and claims undergo human or author checks. Its releases consist of research-licensed audio and include cautions about sensitive affective categories and preliminary annotations.

AI use statement

  • MOSS-Audio generated dense annotations and reasoning traces; EmpathicInsight-Voice and BUD-E-Whisper produced Emolia emotion scores, which are not benchmark ground truth.Every released VoiceNet label is a human rating; human audit bounds the quality of model-generated labels.Source: AI use statement · p. 11

Ethics statement

  • The releases use source audio under open CC-BY or CC-BY-NC research licenses, and the authors caution that sensitive categories and preliminary VoiceNet-Ext annotations are research artefacts, not diagnostic instruments.The statement notes potential privacy implications of affective monitoring and recommends downstream safeguards for deployment.Source: Ethics statement · p. 11

Reproducibility statement

  • Evaluation code, training scripts, manifests, benchmark labels, model checkpoints, corpus data, annotations, and an interactive browser are made available through the listed project and hosting sites.The authors report that scripts regenerated and cross-checked all reported numbers.Source: Reproducibility statement · pp. 11–12

A.1 Comparison to Prior Speech-Emotion Benchmarks

The appendix compares speech-emotion benchmarks, describes the Emolia corpus resources, and lists the emotion taxonomy and talking-style attributes used in VoiceNet.

A.1 Comparison to Prior Speech-Emotion Benchmarks

  • Table 3 compares speech-emotion benchmarks by label-set size, speech type, in-the-wild status, and licensing, with emotion labels distinguished from VoiceNet-Ext talking-style attributes.The table covers benchmarks; MOSS-Audio training corpora are summarized separately.Source: Table 3 · p. 18, §A.1 · p. 18

A.2 EMOLIA and EMOLIA-BALANCED Corpora

  • Emolia annotates the 71.78M-clip Emilia-Large corpus with 40 emotion scores, captions, and speaker-timbre embeddings; Emolia-Balanced is a 5.26M-clip balanced subset.The four MOSS-Audio-annotated corpora and their annotation-call counts are listed separately.Source: §A.2 · p. 18, Table 4 · p. 18

A.3 EMONET-VOICE Taxonomy

  • The 40 EmoneT-Voice emotion categories are listed with associated descriptive terms used during conceptualization and prompting.The taxonomy is adapted from EmoneT-Face.Source: §A.3 · p. 18

A.4 VOICENET-EXT Attribute List

  • VoiceNet-Ext’s 57 attributes span affective dimensions, prosodic delivery, vocal quality, resonance placement, and style descriptors.The schema also includes perceived-speaker and recording-context attributes; several MOSS-Audio outputs are excluded from VoiceNet-Ext.Source: §A.4 · p. 20

A.5 Dataset configuration

The training mixture combines vocal-style captions with retained FCaps captions and simpler Synthetic Vocal Bursts captions. Multilingual In The Wild is released with EMOLIA but excluded from training after an internal ablation.

A.5 Dataset configuration

  • Multilingual In The Wild is annotated and released in the EMOLIA suite but excluded from the training mixture after an internal ablation.The other listed corpora contribute vocal-style, FCaps-style, or simpler captions.Source: §A.5 · p. 20, Table 5 · p. 21

A.6 Evaluation Harness and Loader Audit

The appendix documents the evaluation harness, released per-tag results, annotation reliability, and the construction and agreement properties of the emotion taxonomy. It reports interim human-audit results for model-generated labels and emphasizes limits in interpreting the preliminary attribute subset.

A.6 Evaluation Harness and Loader Audit

  • The harness computes audio–prompt cosine similarities, sweeps thresholds, and uses a per-prompt threshold only when that prompt has at least 10 rows.Otherwise, evaluation uses the global threshold.Source: §A.6 · p. 21

A.7 Per-tag Summary Files

  • Released per-tag files include per-prompt and per-task metrics, similarities, and labels; paired clip-bootstrap intervals use 2,000 samples with per-prompt thresholds held fixed.Those intervals do not propagate VOICENET-EXT annotation uncertainty.Source: §A.7 · p. 21, Table 6 · p. 21

A.8 VOICENET-EMO Per-Emotion ρ

  • Table 7 reports per-emotion Spearman ρ for VOICECLAP, Omni-Embedding bases, and seven general-audio CLAP baselines, with rows ordered by mean ρ.That mean is used only for ordering and is not displayed.Source: Table 7 · p. 22

A.9 VOICENET-EXT Per-Attribute ρ

  • VOICENET-EXT’s mean Spearman ρ is 0.15 for VOICECLAP-LARGE and 0.11 for VOICECLAP-SMALL, while aggregate Fleiss κ is −0.0218 with a 95% interval excluding zero.Per-attribute model values are noisy at about 46 audios per attribute-level prompt.Source: §A.9 · p. 22

A.10 Human Audit of the MOSS-Audio Labels (Interim)

  • Mapped human–MOSS labels match exactly on 0.257 of judgements, but the audit is interim and supports overall statements only, not per-attribute conclusions.The audit covers 97 of 300 planned items; fit judgments were too permissive to serve as the primary measure.Source: §A.10 · p. 23

A.11 Detailed Taxonomy Construction Methodology

  • The taxonomy’s final 40 categories cluster related emotion terms and support context-aware, socially informed multi-label interpretations rather than fixed, single-label classifications.The taxonomy was developed through literature-driven term extraction followed by expert-guided clustering.Source: §A.11 · pp. 23–24

A.12 Per-Emotion Annotator Agreement on VOICENET-EMO

  • Experts are unanimous on 32.7% of VOICENET-EMO items, and per-emotion unanimity has near-zero correlation with VOICECLAP-LARGE ρ (Spearman 0.02).No concentration of model gains appears among higher-agreement emotions.Source: §A.12 · p. 24

A.13 VOICECLAP Training Details

VOICECLAP-SMALL and VOICECLAP-LARGE use distinct contrastive architectures and training configurations. VOICENET-EXT uses binary human judgments of attribute-level matches, but its current reliability results leave the subset preliminary.

A.13 VOICECLAP Training Details

  • VOICECLAP-SMALL is a 110M-parameter dual-tower model trained with SigLIP contrastive loss, while VOICECLAP-LARGE fine-tunes a 7B single-tower model with rank-16 LoRA and symmetric InfoNCE.Both are trained for one epoch; the stated runs take about 0.5 hours for Small and 2 hours for Large on four GH200s.Source: §A.13 · p. 24

A.14 VOICENET-EXT Annotation Coverage and Reliability

  • VOICENET-EXT has near-chance rater agreement: Fleiss’ binary κ is −0.022, and no attribute reaches the prespecified κ ≥0.20 reliability threshold.All 57 attributes are therefore treated as exploratory perceptual probes rather than a finalized benchmark ranking.Source: §A.14 · p. 25

A.15 Additional Results

Additional analyses compare model scores with human references, examine calibration and agreement strata, and diagnose dataset-specific gaps. The paper also contextualizes emotion-ranking correlations using psychological effect-size benchmarks.

A.15 Additional Results

  • VOICECLAP-LARGE reaches 0.7021 bal@pp on VOICENET-EMO versus the 0.572 leave-one-out human reference, indicating closer alignment with aggregate expert consensus, not superior human emotion perception.The model is scored against a majority label, unlike each expert’s comparison against other raters.Source: §A.15 · p. 25
  • Per-prompt calibration raises VOICECLAP-LARGE from bal@0 = 0.580 to bal@pp = 0.702, while VOICECLAP-SMALL changes from bal@0 = 0.659 to bal@pp = 0.675.The Large model’s positive similarity bias accounts for much of its calibration headroom; family ordering is preserved under bal@opt and per-prompt ρ.Source: §A.15 · pp. 25–26
  • None of the four agreement-stratified interactions excludes zero, so the results do not show VoiceCLAP’s edge concentrated in higher-agreement categories.For VOICENET-EXT, “higher” denotes only a within-subset rank, not reliable annotation.Source: §A.15 · p. 26
  • The Large-minus-base gap is −0.0115 on synthetic EmoNet-Voice; on RAVDESS, micro-accuracy falls by 0.0139 while macro F1 rises from 0.285 to 0.330.The results indicate dataset- and taxonomy-specific boundary changes rather than a generic acted-versus-naturalistic explanation.Source: §A.15 · p. 26
  • Fig. 2 plots average bal@pp against model size.The supplied passage describes the plotted quantities but does not state a trend or outcome.Source: §A.15 · p. 26

A.16 Effect Sizes in Psychological Context

  • VOICECLAP-LARGE’s emotion-ranking ρ of 0.372 exceeds the 75th-percentile psychological benchmark correlation of 0.29.The seven general-audio CLAPs each have |ρ| < 0.1 on VOICENET-EMO.Source: §A.16 · pp. 26–27

A.17 Extended Limitations

The limitations bound interpretation of the benchmark: evaluation is embedding-only, VOICENET-EXT is exploratory, and training-corpus coverage and model-generated labels have caveats. The appendix also documents separate expert-rating and binary-confirmation interfaces.

A.17 Extended Limitations

  • All evaluated systems are voice-text embedding models, so the evaluation does not establish performance for assistants, dialogue, response generation, or generative audio-language models.The evaluation uses cosine similarity to rank vocal-attribute descriptions against speech.Source: §A.17 · p. 27
  • The MOSS-label audit is interim: human–MOSS agreement is comparable to human–human agreement, but the 97-of-300-item audit does not establish label correctness.The paired difference is +0.018 with a 95% CI of [−0.053, +0.102], which contains zero.Source: §A.17 · p. 27
  • The released voice corpora inherit source bias, with non-uniform geographic, demographic, and recording-environment coverage; takedown requests will be honoured for two in-house collections.The two Synthetic Vocal Bursts collections comprise 341k clips curated from Creative Commons-licensed YouTube uploads.Source: §A.17 · pp. 27–28
  • VOICENET-EXT has aggregate κ = −0.022, and none of its 57 attributes passes the reliable-core rule, so its model rankings are exploratory.The authors say additional ratings alone are unlikely to stabilise rankings and call for rubric revision with anchor-example rater training.Source: §A.17 · p. 27
  • Bootstrap intervals condition on aggregate labels and fitted thresholds, while four Small runs provide a variance proxy rather than evidence of Large-model seed stability.The authors report that 7B-scale retraining of VOICECLAP-LARGE is infeasible within their compute budget.Source: §A.17 · p. 27

B Annotation Platform Instructions and UI

Related papers

Loading 2609.32016v1…