acceptodds
Under review as a conference paper at ICLR 2027

GASA: A Structured Alignment Benchmark for Text-to-Audio Generation

Abstract

Text-to-audio generation has advanced faster than the methods used to evaluate it. Models are known to miss parts of a prompt, but prevailing measures judge a generated clip against its prompt as a whole and cannot show which parts are missed or where research should focus. We introduce the Generative Audio Structured Alignment (GASA) benchmark. GASA treats prompt fidelity as structured alignment, organizing the properties of a scene description into 25 fields and judging each field with a large audio-language model. Because such judges may answer from text-only priors without listening, we validate judge probes on real caption-audio pairs against a Gaussian-noise control, keeping for each field the probe whose accuracy depends most on the audio. Human raters answering the same probes on generated clips confirm that the judge detects 23 of the 25 fields in generated audio and set the ceiling each field score can attain. Using this validated judge, we evaluate 11 models on 2,500 generation prompts (100 per field). Scores vary far more across fields than across models, and every model leaves several fields near the control. Source identity is strongest, static acoustic properties weakest, and pitch and timbre are rendered more faithfully as changes than as static levels. GASA pinpoints where each model fails, giving researchers specific targets and a measure of progress toward models that realize every part of a prompt. https://gasa-demo.github.io/

Primary area:
datasets and benchmarks
Keywords:
text-to-audio generation, evaluation, benchmark, structured alignment, audio-language models, human evaluation

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_4QQyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GASA, a structured alignment benchmark for text-to-audio generation. It decomposes prompt requirements into 25 fields across seven dimensions and uses controlled minimal prompt pairs to evaluate whether generated audio realizes individual properties. The authors validate field-specific probes using real audio and Gaussian-noise controls, select Qwen3-Omni as the primary judge, and conduct additional human and cross-judge evaluations. The benchmark evaluates 11 generation models on 2,500 prompts, with 23 fields included in the main scoring after excluding two judge-limited fields.

Critical strengths
  • Important problem and actionable evaluation.

    Holistic text–audio similarity can obscure failures in acoustic attributes, temporal relations, and other prompt details. GASA’s field-level evaluation provides a more informative capability profile and clearer targets for model improvement than a single global score.

  • Well-motivated controlled comparisons and judge validation.

    Minimal prompt pairs help isolate responsiveness to individual properties. The real-audio versus Gaussian-noise comparison also explicitly addresses the risk that a judge answers from textual priors rather than listening. Selecting probe formats separately for each field is a thoughtful aspect of the design.

  • Broad experiments and transparent reporting of limitations.

    The paper evaluates multiple generation architectures and includes checks across generation seeds, option orders, and alternative judges. Importantly, the authors identify failures on distance and location through human evaluation and exclude these fields from the main scores rather than attributing the failures to the generators.

Critical weaknesses

Judge reliability remains a concern. It is unclear whether the LLM judge can reliably recognize all evaluated attributes in generated audio. Low scores may reflect the judge failing to recognize an attribute rather than the generator failing to produce it. The documented failures on distance and location reinforce this concern, and further attribute-level validation against human judgments would strengthen the results.

Questions for authors

Further attribute-level validation against human judgments and using additional recent models such as Qwen3.8-Omni-Flash and Gemini 4, would help assess whether the findings are robust to the choice of judge.

2 comments
AuthorsOpenReview ↗

3.1 Judge reliability

"Judge reliability remains a concern. ... further attribute-level validation against human judgments would strengthen the results."

Thank you for your comment. This concern is justified, and that is why the paper has three safeguards specifically against "the judge did not hear it": when selecting the probe for a field, only formats that the judge answers correctly by listening are kept (Section 4.2); for every field, the score obtainable without listening is subtracted (Section 5.2); and fields for which the judge cannot distinguish human-perceivable properties from noise are not included, which is how distance and location were excluded (Section 5.2 and Appendix J). For the other fields, the ceiling is the judge's mean score on the clips that raters confirmed as realizing the field (Section 5.2 and Appendix J).

Since the submission we have collected two further sets of human ratings. The first is an additional round for the eleven fields with the fewest confirmed clips, rated by three of the authors with the same method as the first round. All eleven fields remain separated from their control, and Table 3 now uses the ceilings from all rounds (Appendix J). The second is a re-rating of all 340 clips of the first round by five listeners who are not authors of the paper, with the same questions and interface (Appendix K). The judge agrees with these raters as it does with the authors: its decision matches their majority on 71% of the clips of the 23 scored fields (73% for the authors), and it rates 91% of the clips they confirmed as realized (89% for the authors).

AuthorsOpenReview ↗

3.2 Additional (closed, API-only) judges

"Further attribute-level validation against human judgments and using additional recent models such as Qwen3.8-Omni-Flash and Gemini 4, would help assess whether the findings are robust to the choice of judge."

Thank you for the suggestion. On the stability of the judge, in the paper we re-scored the entire benchmark with two other open-weight judges, Qwen2.5-Omni and AudioFlamingo-Next. The results (in Appendix H of the paper) show the Spearman rank correlations of the model orderings with the main judge are .95 (Qwen2.5-Omni) and .89 (AudioFlamingo-Next), and the three judges all agree on the two strongest and the two weakest models. So, the findings seem robust to the judge choice.

Many thanks for recommending Qwen3.8-Omni-Flash and Gemini 4. Unfortunately, a suitable judge has to be able to return token probabilities because the score used in the paper is the probability of the judge’s answer to a question, that is, the first-token probability it assigns to "yes" and "no" (or to each option letter), and the noise floor, the ceiling, and the normalization are all built on this probability. We checked the official documentation as of November 10: Qwen3.8-Omni-Flash currently has no public weights and can only be called through the Alibaba Cloud API, which, according to its documentation, does not provide token probabilities for the omni series; Gemini 4 is likewise API-only, and models from Gemini 3 onward do not provide token probabilities either. So, currently, these two models cannot be scored under the paper's protocol. Should they provide token probabilities or release their weights in the future, we would be glad to include them in the judge comparison.

Reviewer_gx6FOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Current text-to-audio metrics evaluate generated clips as a whole and hence cannot provide precise feedback across various prompt fidelity dimensions. The authors fill this gap by creating a benchmark which measures performance across 25 fields (e.g., pitch, loudness, distance), covering 7 dimensions (e.g., acoustic, temporal, spatial properties). AI-as-a-judge is used to evaluate the generated audio clips, using the real AudioCaps clips against Gaussian noise to identify comparative performance and retain the question format whose accuracy depends on the audio. Based on this validation method, Qwen3-Omni is selected as the judge from three candidates. Authors also incorporate human raters to listen to generated clips and set a performance ceiling against AI judges. Fields which AI judges are unable to detect in the generated clips are dropped. Authors evaluate 11 models on the resulting 2,300 generation prompts (100 per field). Scores vary more across fields than across models, and every model leaves several fields near the random Gaussian control.

Critical strengths

(1) Strong validation of the benchmarking method: Authors put in appropriate effort to validate the AI judges and audio fields. Each question is scored on Gaussian noise to give a no-listening floor and human raters confirm which generated clips genuinely contain the property to give a ceiling. Fields are included only if the judge separates confirmed clips from the floor and two fields (distance, location) are dropped on this basis. Rescoring with two alternate judges leaves the rankings intact.

(2) Strong motivation and clear practical value: Given continuous progress in text-to-audio generation models, a single alignment score is no longer sufficient to evaluate performance. The current benchmark provides clear benefit and does not show redundancy with existing metrics (e.g., CLAP).

(3) Interesting results that identify concrete directions for future research: Variation across fields exceeds variation across models. All eleven models handle event order, object identity and dynamic properties well but sit near the noise floor on static loudness, duration, pitch and timbre. Since every architecture shares these gaps, they point to the need for broad field-wide improvements.

Critical weaknesses

(1) The taxonomy is the paper's important contribution and it is never validated: The 25 fields are said to be grounded in caption corpus analysis, the AudioSet ontology and perception research, but nothing tests whether they are complete, non-overlapping, or at the right granularity.

(2) No analysis of how the fields relate to one another: Fields are averaged within the seven dimensions and equally weighted in the global aggregate, which assumes each field carries independent information, but no correlation or factor structure is reported. The authors own Shapley analysis in the appendix suggests a few fields carry most of the predictive weight, so the effective dimensionality may be lower than 23, and part of the reported spread across fields may reflect redundancy rather than distinct capabilities.

(3) Small sample for human evaluation: Only five evaluators are used and they are the authors, producing scope for bias in the evaluations.

Questions for authors

(1) Why 25 fields? How were these chosen? (2) How accurate is the Qwen2.5-72B field annotation? Was it checked against human labels on a sample? (3) Have you looked at the correlation between fields? Do they group the way the seven dimensions predict?

4 comments
AuthorsOpenReview ↗

2.1 Where the 25 fields come from and whether they are complete

"The taxonomy is the paper’s important contribution and it is never validated: The 25 fields are said to be grounded in caption corpus analysis, the AudioSet ontology and perception research, but nothing tests whether they are complete, non-overlapping, or at the right granularity."

"Why 25 fields? How were these chosen?"

Thank you for your comments. Section 3 of the paper describes the sources of the 25 fields in one sentence ("These fields are grounded in caption corpus analysis (Appendix B), the AudioSet ontology (Gemmeke et al., 2017), and environmental sound perception research (Gaver, 1993; Guastavino, 2018)''), and Appendix B describes the bottom-up pass based on caption corpora. However, the paper currently does not fully describe the whole process or test the resulting set of fields. Here we provide more detail.

In our bottom-up pass, we analyzed 107,797 captions across the five corpora in Table B.1 and collected the properties people mention when describing a sound, yielding 18 descriptive fields (Appendix B). In our top-down pass, we refined these fields, drawing upon the AudioSet ontology and research on how people hear and describe everyday sounds. We organized these fields along seven dimensions, splitting or merging them so that each field is one property that can be defined on its own and occurs in captions, and then we re-annotated all the captions with the resulting fields, which gives the presence rate of each field in Table B.1. So 25 is the number of fields that resulted from these two steps.

The most direct test of whether the taxonomy is complete is to examine how much of human-authored captions/prompts fit into it. We drew 2,000 captions at random from each of AudioCaps and Clotho and annotated them with the 25 fields using the paper’s annotation model, Qwen2.5-72B-Instruct, which quotes, for each field it marks, the words of the caption associated with it (Appendix B). After removing words that don't provide information, such as a, the, and and, we counted how many belong to one of the 25 fields. In AudioCaps, 99% of these words belong to one of the 25 fields, and in Clotho, 96%. The remaining 1% and 4% don't belong to any field.

The remaining words fall into three categories: 1) adverbs and adverbial phrases that modify the action, i.e., how it unfolds in time (e.g., "throughout", "until it stops") or how fast or how forcefully (e.g., "at high speed", "blowing hard"), 2) words that describe relations among several sources (e.g., "different kinds of birds", "each other"), and 3) context unrelated to the sound itself, such as the time of day or what is being said. In the paper's limitations section, we say that a prompt in this benchmark describes one source, or two sources in a stated relation, so relations among several sources are left to future work. These action modifiers are properties that could also be added as fields in future work. We have added this note to the paper's limitations.

We have added this analysis to the appendix that describes the annotation process (Appendix B, "Coverage of Caption Wording").

AuthorsOpenReview ↗

2.2 Correlation between fields

"No analysis of how the fields relate to one another ... reported spread across fields may reflect redundancy rather than distinct capabilities."

"Have you looked at the correlation between fields? Do they group the way the seven dimensions predict?"

Thank you for your comments. The paper reports the field scores and averages the 23 fields with equal weight, but it does not report how the fields relate to one another. We add that analysis here and clarify the Shapley analysis in Appendix I.

The Shapley values in Appendix I are not about overlap between fields. They measure how much each field score contributes when the field scores are used to predict listeners’ ratings of whole clips on RELATE (Section 4.3). RELATE’s captions mostly state a source and an action and rarely the other fields, so object and action carry most of the weight; this reflects what those captions say and what weight listeners attribute to certain fields, not what the fields have in common. In the benchmark, every field has its own 100 prompts.

In the paper's Table 3, each field has one score per model, and for each pair of fields we computed the Pearson correlation of these 11 values. The mean correlation over the 253 pairs is .18. For the aggregate 7 dimensions, it is .26 within a dimension and .16 between dimensions, so the field scores do not seem strongly correlated. However, these are correlations between the models’ scores; i.e., two fields correlate when the models that do well on one also do well on the other, showing how the models’ capabilities move together, not the acoustic relationship between properties. Testing the acoustical correlation between the fields is a different question that is much harder to measure and show here.

AuthorsOpenReview ↗

2.3 Only five raters, all of which are the authors

"Small sample for human evaluation: Only five evaluators are used and they are the authors, producing scope for bias in the evaluations."

Thank you for your comment. Yes, as stated in the paper, the five raters are authors. However, the human evaluation was strictly blind: the raters could not see the model's name, the judge scores, or how the clips had been sampled. Each clip was judged by three raters with a majority rule, and the pairwise agreement among the raters was 64%. Furthermore, the raters' answers do not directly enter any score; they are used only to confirm whether a generated clip realizes the field.

We nevertheless understand your concern, and so we ran a controlled human evaluation. Five listeners who are not authors re-rated all 340 clips of the first round with the same questions, three raters per clip (Appendix K of the revised paper). On the 23 scored fields, the non-author panel and the author panel reach the same confirmation decision on 79% of the clips, and the non-author raters agree with one another as much as the authors do (pairwise agreement 66% against 64%; Fleiss' kappa .58 against .56).

AuthorsOpenReview ↗

2.4 Accuracy of the automatic field annotation

"How accurate is the Qwen2.5-72B field annotation? Was it checked against human labels on a sample?"

Yes. We evaluated the consistency between model-generated annotations and human annotations on 300 captions. Treating each annotated field as a tag, the model achieved approximately 95% precision and 84% recall against the human annotations, and among the fields identified by both, 97% had matching quoted text spans.

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

The paper proposes a structured alignment benchmark for text-to-audio generation. Instead of evaluating prompt correlation with a single global similarity score, authors decompose audio descriptions into 25 fields across seven dimensions and evaluate them individually using audio-language models as judges. Authors observe that models generally perform better on source identity and certain dynamic properties than on static acoustic attributes and complex inter-source relations.

Critical strengths

It is a well-motivated problem that existing holistic text-to-audio alignment metrics provide limited insight into which specific properties of a prompt are correctly realized. The use of minimal pairs and Gaussian-noise controls is useful for investigating fine-grained alignment and potential language-prior biases in automated judges.

Critical weaknesses

The proposed field-level aggregation does not outperform existing holistic QA metrics on human correlation (Table 2, combined Match + fields + indicators predictor yields a correlation of 0.519 on RELATE, compared with 0.556 for AQAScore.) I understand the paper’s main contribution could be a proposition that is diagnostic interpretability rather than superior overall correlation. However, convincing evidence that the identified weaknesses correspond to real, actionable generation failures is limited, leaving the reader to infer the practical value of the diagnostic scores.

The methodology introduces several heuristic scoring and selection criteria whose specific mathematical forms are not sufficiently motivated.

The paper would benefit from a more principled justification and sensitivity analyses demonstrating that its central findings do not depend strongly on these choices. See questions below.

Questions for authors

Section 4 provides formulation for determining which probe is selected for every field. I understand the intuition to reward audio-dependent performance and penalize text-only leakage. It is unclear to me how was Gain(1-leakage) arrived at, particularly why this exact normalization. Without a clear decision theoretic objective, it is hard to convincingly establish this formulation. Can the authors provide a more principled justification for this criterion? In particular, can they demonstrate that the additional leakage penalty improves agreement with human judgments or leads to more reliable probe selection compared with simpler alternatives such as R-B (real-audio accuracy minus control accuracy)? Or (R - C)/(1 - C) ?

Overall pipeline contains a sequence of ad-hoc choices - Choose the judge using average leakage-scaled gain, choose best probe for each field, estimate Gaussian noise floor, exclude two fields judged unreliably, average remaining 23 equally - which are introduced in a fairly elaborate evaluation instrument without fully establishing that its conclusions are insensitive to the design decisions. The justification for the scoring construction is weaker than I would expect for a benchmark paper that is proposing a new measurement method. Appendix G.3 checks model ranking stability without the ceiling but the paper's main takeaways rely on cross-field comparisons.

Appendix J, some ceilings are based on just 3-5 confirmed examples - Material on 5/20, Duration on 5/20, Pitch 3/10, Timbre 4/10. The upper 95% bootstrap confidence bound over these 3–5 points is then used as a normalization constant in (S - F)/(C - F). I strongly suspect whether these are stable empirical ceilings, particularly for material, duration, loudness, pitch, and timbre. And how much the cross-field scores shift under point-estimate ceilings or larger human samples.

5 comments
AuthorsOpenReview ↗

1.1 Practical value of diagnostic scores

"The proposed field-level aggregation does not outperform existing holistic QA metrics on human correlation ... leaving the reader to infer the practical value of the diagnostic scores."

Thank you for the feedback. This is an important point, so we address it in a few ways below.

Contextualizing and fixing the AQAScore comparison

You are right that the goal of the paper is diagnostic interpretability rather than superior overall correlation. However, your comment also indicates that we did not clearly contextualize some of the results in Table 2, particularly the Match+fields+indicators 0.519 and the AQAScore 0.556 correlation on RELATE. These two numbers were not obtained under the same conditions. 0.556 is the value reported in the AQAScore paper, with Qwen2.5-Omni-7B as the judge and their own question wording, but that paper does not state which RELATE clips that it scored. When we tried to recreate their results (still with Qwen2.5-Omni-7B), AQAScore had a 0.467 Spearman correlation on the full RELATE test split (1311 real and generated clips), and a 0.523 Spearman correlation with the generated subset of the RELATE test split (874 generated clips). Thus, we suspect their reported score was with the generated subset. On the other hand, the 0.519 value we reported for the Match+fields+indicators was on the full RELATE test split (and used a Qwen3-Omni judge). So the two numbers cannot be directly compared since the clip setes are not the same.

To remedy this issue and fairly compare AQAScore to our approach, we scored the test set with the AQAScore on all pairs and on the generated clips alone, and then added our 23 field scores to each whole-caption score with the same random-forest fit as the last row of Table 2 (Appendix I). In the AQAScore rows the whole-caption score comes from Qwen2.5-Omni-7B, the judge validated in their paper, and the field scores from Qwen3-Omni, the judge validated in ours (Table R1).

Table R1. Spearman correlation with the listeners’ relevance ratings on the RELATE test set, for each whole-caption score alone and with our 23 field scores added, with the 95% interval of the paired gain. The gain is relative to the raw whole-caption score; Appendix I of the revised paper reports it relative to the fitted Match row of Table 2 (.478 to .519, +.042).

Whole-caption scoreClipsAloneWith our fieldsGain (95% CI)
Ours (Qwen3-Omni, our wording)all 1,311.480.519+.039 [+.012, +.067]
Ours (Qwen3-Omni, our wording)874 generated.469.509+.040 [+.009, +.072]
AQAScore (Qwen2.5-Omni-7B, their wording)all 1,311.467.533+.065 [+.038, +.094]
AQAScore (Qwen2.5-Omni-7B, their wording)874 generated.523.550+.027 [-.001, +.056]

With our field scores added, the correlation to human ratings increases in every row, so with either whole-caption score, ours or AQAScore’s, the field scores add information relevant to human ratings that the whole-caption question does not capture. We have clarified this in the revised paper: Table 2 now reports the AQAScore value we recomputed on the full test split (0.467), and its caption and Section 4.3 give the 0.556 reported in the AQAScore paper. We have also added Table R1 to the appendix, as Table M.1 in Appendix M.

AuthorsOpenReview ↗

1.1 Practical value of diagnostic scores (continued)

Demonstrating CLAP vs GASA judge failures

However, this alone doesn't clearly communicate the practical value of our field scores over holistic metrics like CLAP for diagnosing real generation failures. To demonstrate this, we performed an additional analysis in which we tested our judge on audio where a property's presence or absence is known, so any miss is a judge error. We did this for both real recordings and generated clips, and compared the judge with CLAP on the same clips.

For the real audio, we used the Contrast questions of the validation study in Section 4.1 from the AudioCaps validation and test splits. Each question takes a recording aa and its caption cc, and pairs the caption with a copy mm in which the tested field is replaced by its opposite value, so the original caption cc is the correct answer. The GASA judge is correct when P(c∣a)>P(m∣a)P(c | a) > P(m | a), where PP is the probability it assigns to each of the two captions, and CLAP is correct when S(a,c)>S(a,m)S(a, c) > S(a, m), where SS is its similarity between the audio and a text. The GASA judge is correct on 88% of the questions and CLAP on 71%.

For the generated audio, we used the 215 benchmark clips that raters confirmed as realizing the tested field in the first round of the human evaluation (Appendix J). Each clip aa was generated from one prompt pp of a minimal pair, and since the raters confirmed that the clip realizes the field, pp is the correct answer and the other prompt of the pair, qq, is the wrong one. CLAP is correct when it finds the clip more similar to its own prompt, S(a,p)>S(a,q)S(a, p) > S(a, q). The GASA judge answers the benchmark's question for that field, in which one option, kk, agrees with pp. It is correct when it gives kk a higher probability than every other option oo, P(k∣a)>P(o∣a)P(k | a) > P(o | a). The GASA judge is correct on 89% of the clips and CLAP on 65%. So the judge is generally reliable (more reliable than CLAP), and the low field scores we report come mainly from properties that the models did not realize.

AuthorsOpenReview ↗

1.1 Practical value of diagnostic scores (continued)

We also picked examples to show what such a failure looks like, next to clips in which the same fields succeed (Table R2). From the benchmark, we took clips with a high CLAP score within their model and a low field score (clips 1 to 3). From the generated clips of the RELATE test set, which the judge scores as in Section 4.3, we took clips on which the judge answers the object and action questions correctly and the question for one other field fails (clips 4 to 6), and clips on which the questions for three of the same four fields succeed (clips 7 to 9). We listened to each clip: the failures are described in the table, and clips 7 to 9 match their captions. CLAP is .28 to .61 for the failures and .41 to .48 for the successes, and Match is .64 to 1.00 against .98 to 1.00, so for both scores the range of the successes lies inside that of the failures, whereas the question for the tested field is at most .03 against at least .97. We have added the spectrograms and loudness curves of these clips, with the events marked, in Appendix L of the revised paper (Table L.1; Figures L.1 to L.3).

Table R2. Example clips: one field fails in clips 1 to 6, and three of these fields succeed in clips 7 to 9. The first part gives the text of each clip and how the audio differs from it, and the second the scores. CLAP is the cosine similarity between the text and the clip, computed as in Appendix F; its median over the generated clips of the RELATE test set is .42. Match and the field columns give the judge's probability of yes, the correct answer, as the mean over the questions when the text names several sources or events. Listeners is the mean relevance rating of the clip in RELATE (0 to 10).

ClipPrompt or captionHow the audio differs from the prompt / caption
1A man whistles, then a crowd applauds.Applause from the start.
2A goat bleats, then a man laughs.The man laughs first.
3A man laughs, then a crowd cheers.The crowd cheers first.
4A loud engine roars followed by four loud beeps.Eight beeps, not four.
5A man speaks while operating a power toolThe tool starts after he stops.
6A door repeatedly slamming as a group of men and women talkThe door slams once.
7A man speaks followed by a toilet flushNo difference
8Water running softly in the background followed by several shrill beeps and a man speakingNo difference
9A clock ticking repeatedlyNo difference
ClipModelCLAPMatchObjectActionTested fieldListeners
1Tango 2.528.96.90.98temporal sequence .01-
2AudioX.609.91.991.00temporal sequence .00-
3Stable Audio Open.557.67.95.99temporal sequence .00-
4Tango.560.991.00.81count .035.8
5Tango 2.4231.001.00.96temporal concurrency .016.5
6Tango.275.64.93.99repetition .035.0
7Tango.4741.001.00.99temporal sequence 1.009.4
8Tango 2.4811.00.99.99count .978.8
9Tango 2.406.98.97.99repetition .989.8

The captions of clips 4, 6 and 8 state further fields, which the judge scores at .89 (loudness) and 1.00 (temporal sequence) for clip 4, at .93 (human specifics) and .41 (temporal concurrency) for clip 6, and at .84 (loudness), .97 (timbre), .97 (temporal sequence) and .99 (foreground/background) for clip 8.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.