PaperlayerView exact revision on arXiv

Source-linked AI summary

VocalGrad: Evaluating Acoustic Perception in Audio Language Models

Koki Ryu, Hitomi Yanaka

arXiv:2610.23745v1cs.SDcs.CL

TL;DR

Existing benchmarks make it difficult to distinguish acoustic-perception failures from downstream reasoning or semantic integration. VocalGrad isolates temporal acoustic changes in a binary task and combines text evaluation with linear probing. Humans perform strongly while LALMs remain near chance, with probing revealing distinct representational bottlenecks.

  • Problem

    Existing benchmarks do not clearly separate acoustic perception from higher-level reasoning, limiting diagnosis of LALMs’ difficulties using acoustic information.

    Source: §1 · pp. 1–2

  • Method

    VocalGrad applies a temporal change to one acoustic feature and evaluates increase-versus-decrease judgments, alongside linear probing of internal representations.

    Source: §1 · p. 2, §6 · p. 10

  • Results

    Human annotators perform near perfectly while LALMs remain near chance; probing finds that information is either underused by the language-model head or becomes less linearly decodable across layers.

    Source: §1 · p. 2, §6 · p. 10

  • Takeaways & Limitations

    Text-based evaluation alone cannot reveal whether acoustic information is retained but unused or progressively degraded across model layers.

    Source: §1 · p. 2, §6 · p. 10

    The prompting setup requires models to associate each named transformation with its corresponding acoustic phenomenon.

    Source: §3.4 · p. 5

Abstract

from arXiv · show

Recent large audio language models (LALMs) have extended their capabilities beyond lexical speech understanding to tasks requiring paralinguistic and non-linguistic audio understanding, such as emotion recognition. However, these models still struggle to reliably utilize acoustic information, particularly when it co-occurs with spoken linguistic content. Existing benchmarks do not clearly distinguish acoustic perception from higher-level reasoning, making it difficult to identify the source of these failures. To address this issue, we introduce VocalGrad, a benchmark designed to evaluate LALMs' ability to perceive low-level acoustic variations. VocalGrad formulates this problem as a simple binary classification of temporal change direction, enabling evaluation largely independent of higher-level reasoning or semantic understanding. Our benchmark results reveal a substantial gap in acoustic perception between human annotators and state-of-the-art LALMs: while human annotators achieve high accuracy, LALMs perform at near-chance levels. We further conduct linear probing to identify structural bottlenecks in these models. Our analysis reveals two distinct failure modes: for some categories, task-relevant information is preserved in internal representations but not utilized by the language model head; for others, the information becomes progressively less linearly decodable across model layers. Overall, our results reveal substantial limitations of LALMs in modeling acoustic variation under controlled settings and expose representation differences not observable from text outputs.

1 Introduction

Although LALMs handle paralinguistic and non-linguistic tasks, existing evaluations do not isolate acoustic perception from semantic integration or reasoning. VocalGrad tests temporal change direction with a controlled binary task, revealing near-chance LALM performance and distinct representation bottlenecks.

1 Introduction

  • Prior findings show LALMs struggle to use acoustic cues alongside lexical content, but do not locate whether failures arise in perception or downstream semantic integration [11] [12].Existing primitive-cue benchmarks can also require extra reasoning, making errors hard to attribute.Source: §1 · pp. 1–2
  • VocalGrad changes one acoustic feature over time and asks whether it increases or decreases, largely isolating perception from semantic understanding and higher-level reasoning.Its design spans speaker-intrinsic and environment-induced factors.Source: §1 · p. 2, Figure 1 · p. 2
  • VocalGrad comprises 21,600 clips, and five state-of-the-art LALMs perform near chance while human annotators find the tasks easy.The benchmark systematically isolates acoustic-feature perception.Source: §1 · p. 2
  • Probing identifies two failure modes: some transformations remain represented but are not used by the language-model head, while others become less linearly decodable across decoder layers.Fine-tuning specific acoustic transformations also induces cross-category changes in representations and performance.Source: §1 · p. 2

2 Related Work

Prior acoustic-perception benchmarks often mix perception with additional reasoning, while earlier representation analyses focus on different attributes or model components. VocalGrad isolates temporal changes in low-level features and traces task-relevant information across the LALM pipeline.

2.1 Benchmarks for Acoustic Perception

  • Existing benchmarks test acoustic properties, but often require cross-clip comparisons, reranking, timestamp answers, or high-level attributes, complicating direct assessment of perception.Prior work reports difficulties using prosodic, environmental, and other acoustic information [11] [12] [14].Source: §2.1 · p. 3
  • VocalGrad instead changes one acoustic factor and asks for its temporal direction, making evaluation largely independent of downstream reasoning.This controlled task addresses the attribution problem in earlier benchmarks.Source: §2.1 · p. 3

2.2 Analysis of Internal Representations in Audio Models

  • Prior representation studies examine acoustic and semantic encoding in self-supervised speech models [20] [21] [22] and higher-level attributes in LALMs [23] [25].These studies include layer-wise probing and analyses of information transfer to language components.Source: §2.2 · p. 3
  • VocalGrad tracks low-level temporal-change information from pre-LLM audio representations through decoder layers to text-prediction representations, distinguishing information degradation from failure to use retained information.This extends analysis beyond representations within the language decoder alone.Source: §2.2 · p. 3

3 Dataset Construction

VocalGrad uses balanced binary direction labels and shared speech sources to create a controlled benchmark spanning nine acoustic transformations. Its common augmentation structure supports comparisons across transformations, intensity levels, and temporal trajectories.

3.1 Overview

  • The dataset contains 21,600 clips across nine transformation categories, with balanced increasing and decreasing samples, three intensity tiers, and four temporal trajectories.The categories cover three speaker-intrinsic and six environment-induced transformations.Source: §3.1 · p. 4

3.2 Common Pipelines

  • The common pipeline selects 100 VCTK source clips and applies each transformation across direction, intensity, and trajectory combinations, yielding 24 augmented samples per source clip.The same sources are used across categories; segments are transformed with smooth transitions where applicable.Source: §3.2 · p. 4

3.3 Category-Specific Augmentations

  • The nine category-specific transformations span speech rate, pitch and gain changes, as well as distortion, roughness, noise, echo, clarity, and vibration effects.Examples include pitch shifting with librosa and adding Gaussian background noise.Source: §3.3 · pp. 4–5

3.4 Prompting Format

  • A shared prompt asks whether the named attribute increases or decreases over time, so correct responses require associating each attribute name with its acoustic phenomenon.Human annotators reliably make this association and answer correctly.Source: §3.4 · p. 5

4 Benchmarking

On VocalGrad, humans substantially outperform the evaluated LALMs, which remain near chance. Removing linguistic content improves some results but does not eliminate the models’ difficulty with simple acoustic changes.

4.1 Evaluation Settings

  • The evaluation covers commercial and open-source LALMs, uses text-output accuracy, and compares model scores with annotations from six English-speaking annotators.Human annotations use sampled clips and at least three annotators per clip.Source: §4.1 · p. 5

4.2 Results

  • 94.4% average human accuracy contrasts with near-chance LALM performance; even the best model, Kimi-Audio, achieves only 51.8% overall.The results are reported across categories, and alternative prompts leave performance near chance.Source: §4.2 · pp. 5–6, Table 2 · p. 6

4.3 Ablation: Removing Linguistic Content

  • On sine-wave subsets for four transformations, some models improve, but none exceed 70% accuracy in any category, so removing linguistic content does not resolve the challenge.The controlled ablation uses sine waves lacking speech-specific temporal or spectral structure.Source: §4.3 · p. 6

5 Linear Probing

Linear probing tracks where VocalGrad-relevant acoustic information remains accessible across model representations and decoder layers. The results distinguish information that persists but is poorly read out from information that becomes less linearly decodable, while LoRA fine-tuning changes both task performance and representations.

5.1 Method Overview

  • Linear probes classify increase versus decrease from representations across audio-processing stages and decoder layers to track changes in linear accessibility.The analysis includes audio-token-aligned representations and output-prediction-position representations.Source: §5.1 · p. 7

5.2 Experimental Setup

  • Probe training uses 12,000 samples per transformation category, generated from 500 base clips by 57 VCTK speakers excluded from the test set.Source: §5.2 · p. 7

5.3 Results

  • Late-LLM probing accuracy falls below Pre-LLM levels in more than half of categories, but remains near-perfect for Volume, Background Noise, and Audio Roughness.Text-prediction representations aggregate relevant information around layer 5; it may persist or attenuate in deeper layers.Source: §5.3 · p. 8, Figure 2 · p. 8, Figure 3 · p. 8

5.4 LoRA Fine-tuning

  • LM-Head fine-tuning reaches 88–99% accuracy on Volume, Background Noise, and Audio Roughness, while All-Linear fine-tuning achieves near-perfect accuracy in most categories.The results support a language-model-head bottleneck for the first three categories and suggest earlier layers matter for remaining gaps.Source: §5.4 · p. 9
  • Single-category fine-tuning often improves related categories, while some cross-category accuracies fall significantly below chance, reflecting systematically inverted decision boundaries.Audio Distortion fine-tuning improves Audio Roughness and Background Noise.Source: §5.4 · p. 9, Figure 4 · p. 9
  • Fine-tuning can improve probing accuracy on unseen categories, including near-perfect final-layer Speaking Speed accuracy after Voice-Pitch tuning, without always improving end-to-end accuracy.The authors report representational changes that extend across categories even when not reflected in textual predictions.Source: Figure 4 · p. 9, §5.4 · p. 10

6 Conclusion

VocalGrad evaluates acoustic perception in LALMs using text-based classification and linear probing. The analysis finds that acoustic changes humans readily perceive remain challenging for current models, with different representation bottlenecks across categories.

6 Conclusion

  • For Volume, information remains in final-layer representations while the language-model head is the primary bottleneck; for Voice Pitch and Speaking Speed, information becomes less linearly decodable in deeper layers.Source: §6 · p. 10
  • LoRA experiments show that head-only updates can substantially improve performance for categories with preserved representations, while updating all linear layers improves task performance and yields cross-category effects and enhanced separability.Source: §6 · p. 10
  • Future work will investigate whether improved acoustic perception leads to consistent, generalizable gains on emotion recognition and speech-event understanding.Source: §6 · p. 10

A Limitations

The limitations bound VocalGrad’s conclusions to controlled, isolated changes and leave aspects of probing and category differences unresolved. The models use varied acoustic representations and training designs, but their ability to capture primitive acoustic variation remains unclear.

A Limitations

  • VocalGrad does not establish that findings on isolated temporal changes generalize to naturally occurring acoustic variation, where factors may co-vary.The authors identify evaluation on naturally occurring, co-varying changes as complementary future work.Source: A Limitations · p. 15
  • The main linear probes average representations over 16 fixed temporal segments, so finer-grained or nonlinear analyses may recover information this setup misses.The authors also call for comparisons across temporal resolutions and models better suited to temporal sequences, such as LSTMs.Source: A Limitations · p. 15
  • The authors do not explain why temporal changes remain decodable in some categories but not others; the proposed cue-duration account remains unvalidated.They suggest GeMAPS as a route to clearer analysis.Source: A Limitations · p. 15

B.1 Acoustic Modeling in LALMs

  • Audio Flamingo 3 [7] uses AF-Whisper, initialized from Whisper large-v3 and trained with next-token audio captioning that combines transcripts, ambient sounds, and music metadata.Source: §B.1 · p. 15
  • Step-Audio 2 [6] pretrains its audio encoder on speech and audio tasks, then freezes it and projects its representations into the language model with an adaptor.Source: §B.1 · p. 15
  • Kimi-Audio [5] combines ASR-derived semantic tokens with continuous Whisper-based acoustic features, processed through shared transformer layers and separate text and audio heads.Its training objectives include unimodal modeling, audio-text mapping, and interleaved audio-text prediction.Source: §B.1 · p. 15
  • MiMo-Audio [31] autoregressively models high-fidelity acoustic tokens from a dedicated tokenizer trained with reconstruction and audio-to-text objectives.Its tokens are grouped into patches and processed over interleaved text and audio sequences; the design aims to preserve acoustic as well as semantic information.Source: §B.1 · pp. 15–16
  • Despite these varied modeling strategies, whether the models reliably capture primitive acoustic variations remains unclear.Source: §B.1 · p. 16

B.2 Detailed Description of Temporal Trajectories for Data Augmentation

The appendix specifies VocalGrad’s temporal augmentation, category-specific settings, annotation and runtime procedures, and trajectory-wise results. Human trajectory results are confounded by a sampling error, while models remain near chance across trajectories.

B.2 Detailed Description of Temporal Trajectories for Data Augmentation

  • The augmentation section defines trajectory-specific progress functions over normalized speech time and applies four trajectories, including quadratic variants with a flat start or end.Transformations update at each sample for Echo and Voice Vibration, and in quantized 50-ms windows for the other seven categories; Audio Distortion and Audio Roughness reverse the strength-to-label mapping.Source: §B.2 · p. 16

B.3 Hyperparameters for Dataset Generation

  • Speaking Speed, Voice Pitch, and Volume vary time-stretching, pitch shift, or gain across clips, with high, medium, and low intensity settings.The listed maxima are Δmax = (0.4, 0.3, 0.2) for speed, (3, 2, 1) semitones for pitch, and (18, 12, 6) dB for volume.Source: §B.3 · p. 16
  • The study simulates clarity with low-pass filtering, vibration with 0–20 Hz amplitude modulation, distortion by downsampling 48 kHz audio, roughness by reducing default 16-bit depth, and noise with additive Gaussian noise.High, medium, and low tiers use clarity cutoffs of 1.2, 2.5, and 5.0 kHz; vibration Amax values of 0.9, 0.75, and 0.6; roughness depths of 4, 6, and 8 bits; and noise rmax values of 0.12, 0.07, and 0.03 times speech RMS.Source: §B.3 · pp. 16–17

B.4 Human Annotation

  • Six annotators answered up to 50 binary questions per assigned category, but a sampling error coupled trajectory with direction, limiting trajectory-wise interpretation of human accuracy.The annotation tool allowed multiple replays, and participants could consult resources for attribute terms.Source: §B.4 · p. 17

B.5 Runtime Configuration of Models

  • Runtime settings include bfloat16 for most models, deterministic greedy decoding for other open-source models where available, and one-epoch LoRA comparisons using 10,800 training samples.Gemini 3 Flash used low thinking level and temperature 1.0; probing used a linear classifier trained with AdamW.Source: §B.5 · pp. 18–19

C.1 Per-Temporal-Trajectory Analysis of Benchmarking Results

  • All models remained near chance across trajectories, with no clear systematic trajectory effect; humans performed best on jump, but trajectory and direction were coupled in their subset.Table 5 reports the trajectory-wise breakdown, but the sampling error means human results cannot separate trajectory effects from direction effects.Source: §C.1 · p. 19, Table 5 · p. 19

C.2 Per-Intensity-Tier Analysis of Benchmarking Results

Tier-wise results indicate that weaker pitch changes are harder for human annotators, while stronger background-noise changes modestly help LALMs without moving them far from chance.

C.2 Per-Intensity-Tier Analysis of Benchmarking Results

  • The tier-wise analysis examines accuracy for Background Noise and Voice Pitch across transformation intensities.Source: §C.2 · p. 19
  • For Voice Pitch, human accuracy decreases as transformation intensity weakens, indicating that lower-intensity changes are harder to perceive.Source: §C.2 · p. 19
  • For Background Noise, humans are near-perfect across tiers, while LALMs improve modestly at higher intensities but remain close to chance overall.Source: §C.2 · p. 20

C.3 Analysis of Parsing Failure in Benchmarking

The analysis uses rule-based parsing to reduce errors from non-exact model outputs. Exact-match evaluation especially lowers Gemini 3 Flash accuracy, whose non-exact responses vary substantially by category.

C.3 Analysis of Parsing Failure in Benchmarking

  • Outputs count as correct only when case-insensitive whole-word matching finds exactly one direction label; all other outputs are incorrect.Accuracy is computed over evaluated clips.Source: §C.3 · p. 20
  • Rule-based parsing raises Gemini 3 Flash’s target-label match rate from 83.9% under exact matching to 99.1%, so it is used for fair model comparison.Open-source models have negligible parse error rates.Source: §C.3 · p. 20
  • Under exact-match scoring, open-source model performance remains largely unchanged, while Gemini 3 Flash accuracy drops across several categories.Source: §C.3 · p. 20
  • Gemini 3 Flash’s non-exact outputs include clear full-sentence answers, incomplete responses, and refusals or uncertainty; rule-based parsing recovers the clear answers.Source: §C.3 · p. 20
  • Gemini 3 Flash averages a 16.1% non-exact-match rate, reaching 67.1% for Audio Roughness, with rates varying substantially across categories.The authors suggest response behavior may depend on acoustic properties or the queried attribute.Source: §C.3 · p. 20, Table 8 · p. 21

C.4 Statistical Significance of the Benchmarking Results

The confidence-interval analysis finds model performance statistically indistinguishable from chance for most categories, while human performance remains substantially higher. One model-category result falls significantly below chance.

C.4 Statistical Significance of the Benchmarking Results

  • The study estimates category- and model-specific intervals by simulating predictions at the original sample sizes and taking the 2.5th and 97.5th percentiles.The original sample sizes were n = 150 for humans and n = 2400 for LALMs.Source: §C.4 · p. 21
  • Most model confidence intervals overlap the 50% baseline, while human intervals show a substantial performance gap from models.Intervals are 95% confidence intervals estimated from 10,000 binomial simulations.Source: §C.4 · p. 21
  • MiMo-Audio falls significantly below 50% on Audio Roughness, possibly perceiving the feature but inconsistently mapping it to the correct label.The passage notes that parsing errors do not significantly affect this category.Source: §C.4 · p. 21

C.5 Failure Analysis with Output Distribution

Output distributions reveal strong response biases that help explain near-chance performance. Biases can be global or category-specific, and probing offers one possible account of how they relate to acoustic information.

C.5 Failure Analysis with Output Distribution

  • Step-Audio 2 mini and Audio Flamingo 3 show extreme single-response biases across categories, explaining their near-chance performance.Table 9 reports the share of “increase” responses on VocalGrad and original source clips.Source: §C.5 · p. 22
  • Biases also vary by category: Gemini 3 Flash predicts “increase” for 2.1% of Echo samples, versus 90.9% for MiMo-Audio.The biases become more pronounced on original VCTK source clips.Source: §C.5 · p. 22
  • The authors suggest that when acoustic evidence is unreliable, models default to biased or degenerate responses; probing finds some relevant cues preserved but underused by the language-model head.The proposed explanation is that the head may fall back on prior biases or unstable decision boundaries.Source: §C.5 · p. 22

C.6 Prompt Sensitivity of Benchmarking Results

Prompt and task variations generally leave VocalGrad performance near chance, although wording can shift answer biases. Results across source datasets and alternative formats show similar end-to-end difficulties, while positional probing finds that acoustic information varies across token positions.

C.6 Prompt Sensitivity of Benchmarking Results

  • Swap-order and audio-reference prompts leave open-source models near chance, with no meaningful improvement over the base prompt.The audio-reference version explicitly directs attention to the input audio.Source: §C.6 · p. 23, Table 11 · p. 23
  • Cross-attribute queries produce above- or below-chance Kimi-Audio accuracy on some categories, suggesting performance depends on matching acoustic signals to attribute words.The analysis is limited for Speaking Speed and Voice Pitch because VocalGrad lacks similar-meaning alternative attribute words.Source: §C.6 · p. 23, Figure 8 · p. 23
  • Across nine category-specific prompts, accuracy remains 50.0%–53.2% while wording substantially shifts increase-response rates, so phrasing changes do not reliably improve acoustic-direction sensitivity.For Voice Pitch, increase responses range from 0.1% to 40.1%; for Speaking Speed they range from 0.9% to 26.4%, while the highest accuracy is 53.2%.Source: §C.6 · p. 24, Table 13 · p. 24

C.7 Few-shot Prompting

  • Few-shot prompting with four same-category examples does not substantially improve performance, which remains close to chance overall.A slight improvement is observed for Step-Audio 2 mini.Source: §C.7 · pp. 24–25, Table 14 · p. 25

C.8 Benchmarking Results with Different Source Datasets

  • Kimi-Audio remains near chance across source datasets, averaging 50.6% on Loquacious Set and 52.9% on Common Voice.The Common Voice clips come from spontaneous speech in five low-resource languages.Source: §C.8 · p. 25, Table 15 · p. 25

C.9 Benchmarking Results with Different Problem Formats

  • Alternative change-detection and ternary formats do not resolve Kimi-Audio’s failure: ternary macro-F1 is 21.6%–27.6%, below 33.3% chance, amid strong answer dominance.The source clips used as “constant” may contain natural acoustic fluctuations, so that label means no systematically imposed VocalGrad trajectory.Source: §C.9 · p. 26, Table 16 · p. 26

C.10 Linear Probing Across Different Token Positions

  • Position-wise probes find less decodable information at middle audio positions a2 and a3 than at a1 and a4; under Audio→Query ordering, information concentrates at final query and output positions in deeper layers.The analysis uses the final token of each of four temporal audio segments and the output-prediction position; swapping context order leaves end-to-end accuracy near chance at 50.6%.Source: §C.10 · p. 27, Table 18 · p. 27, Table 19 · p. 27

C.11 Fine-tuning Results in All Target Models

This section presents VocalGrad fine-tuning results across benchmark categories, including category alignment and downstream emotion recognition, and covers broader-impact and research-practice disclosures. These disclosures address societal benefits and risks, dataset documentation and licensing, human annotation, LLM use, and limitations.

C.11 Fine-tuning Results in All Target Models

  • All-Linear fine-tuning substantially improves performance across most categories in all four open-source models, while LM-Head results track late-layer probing trends.MiMo-Audio’s relatively higher Voice Clarity probing accuracy is reflected in its LM-Head fine-tuning results; the table reports performance for all fine-tuned models.Source: §C.11 · pp. 27–28, Table 20 · p. 28

C.12 Representation Alignment Across Categories After Fine-Tuning

  • At layer 28, mean absolute cosine similarity rises from 0.205 in base Kimi-Audio to 0.729 after Volume tuning and 0.682 after Voice-Pitch tuning.Mean signed similarity remains near zero because category directions show both large positive and negative alignments, becoming approximately parallel or antiparallel; this alignment also includes categories not used for tuning.Source: §C.12 · pp. 28–29, Table 21 · p. 28, Figure 9 · p. 29
  • Speaking Speed remains weakly aligned with several categories even after fine-tuning, despite improved probing after Voice-Pitch tuning, so shared directions alone do not explain cross-category effects.Fine-tuning may also change category-specific representations or their linear separability, which this direction-based analysis does not capture.Source: §C.12 · p. 29

C.13 Transfer of VocalGrad Fine-Tuning to Emotion Recognition

  • Voice-Pitch tuning shifts MELD predictions toward Joy, with 68.9% Joy recall and 25.8% of samples called Neutral by the base model instead predicted as Joy.Several other class results decline, and the findings do not establish beneficial transfer to downstream paralinguistic tasks; broader multi-attribute training and systematic natural-speech evaluation are needed.Source: §C.13 · pp. 29–30
  • MELD macro-F1 falls from 41.3 for the base model to 37.7 after Volume fine-tuning and 34.2 after Voice Pitch fine-tuning.Volume fine-tuning raises Neutral F1 from 66.4 to 67.3, but Voice Pitch fine-tuning shifts prediction behavior, including raising Joy recall to 68.9%; neither condition improves overall macro-F1.Source: §C.13 · p. 29, Table 22 · p. 30

E Licenses of the Existing Assets

  • The models are released for research use: Kimi-Audio, MiMo-Audio, and Step-Audio 2 mini under MIT, and Audio Flamingo 3 for non-commercial research.VocalGrad inherits VCTK’s CC BY 4.0 license; Gemini 3 Flash use follows the Gemini API terms.Source: E Licenses of the Existing Assets · p. 30

F Broader Impacts

  • VocalGrad uses controlled transformations of public VCTK speech and adds no personal or sensitive information beyond the original recordings, though real speech still poses potential risks.Source: F Broader Impacts · p. 31

2. Limitations

  • The paper identifies limited generalization to naturally occurring co-varying acoustic changes, fixed temporal resolution in linear probing, and limited analysis of category-level information retention.AI-assisted tools were used for writing refinement and coding assistance, not core methods or experimental design.Source: G AI Assistance Usage · p. 31, §2 · p. 32

10. Broader impacts

  • The authors discuss benefits for LALM evaluation and potential speech-data and privacy risks, while stating that the public-data augmentations add minimal misuse risk.They also describe mitigation strategies and report that the study collected no personal or sensitive participant data and reports only aggregated results.Source: §10 · pp. 36, 38
  • The authors state that VocalGrad is publicly released with usage and reproduction instructions, metadata, and licensing information, and that existing assets are credited and used under their terms.They report using VCTK, Loquacious Set, Common Voice, MELD, and several audio language models.Source: §10 · p. 37
  • The paper reports providing annotation instructions, an interface screenshot, and compensation information in Appendix B.4.Participants were informed about the study and participated voluntarily with compensation offered; the authors describe the binary annotation task as low risk.Source: §10 · p. 38
  • The authors state that LLM-based tools supported coding and writing but did not contribute to the core methodology, experimental design, or results.Source: §10 · p. 38

Related papers

Loading 2610.23745v1…