Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Domain Adaptation Method and Modality Gap Impact in Audio-Text Models for Prototypical Sound Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Subtracting a scaled background profile restores zero-shot sound classification accuracy when background noise is present.

desk verdict A simple background-profile subtraction clearly helps zero-shot audio-text classification, but the headline audio-based gains may be an oracle artifact from reusing the exact test background recordings. read the letter →

arxiv 2506.04376 v1 pith:OBDBYYKE submitted 2025-06-04 cs.SD cs.AIeess.AS

Emiliano Acevedo , Martín Rocamora , Magdalena Fuentes This is my paper
classification cs.SDcs.AIeess.AS
keywords audio-textmodelszero-shotsoundclassificationdomainadaptationmodalitygapprototypicalbackgroundsoundscapeurbancosinesimilarityprofile
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper shows that zero-shot sound classification with audio-text models collapses when a background soundscape is mixed with the target sound, and that the damage is controlled by signal-to-noise ratio rather than by which background is present. It proposes a domain adaptation that builds a profile of the test audio's similarity to all class prototypes, builds a separate profile for the background alone, and subtracts a scaled version of the background profile from the test profile. Across six urban backgrounds and three SNR levels, this subtraction improves accuracy for both text-anchor prototypes and TGAP audio prototypes, with audio-based background profiles giving the largest gains. The paper also argues that narrowing the audio-text modality gap, for instance by using unlabeled audio to build prototypes, consistently improves classification. If the method holds, zero-shot classifiers can be made more robust to real-world acoustic scenes without retraining.

What carries the argument

The method's central object is the profile: a vector whose entries are cosine similarities between an audio embedding and each class prototype in the shared audio-text space. Classification normally takes the largest entry; the paper instead computes the test audio profile $P_s$ and a background profile $P_b$ (from textual background prompts or from a background audio recording) and uses $P_f = P_s - P_b \times \tau$ as the refined profile. Equation (1) is the engine of the method: it treats the background as a set of individual sound sources whose similarity contributions add to the foreground's, and the scalar $\tau$ controls how much of that contribution is written off. The paper sets $\tau=0.2$ for text-based adaptation and $\tau=0.7$ for audio-based adaptation.

What would settle it

Find a background-foreground pair where the linear subtraction leaves a large systematic residual, for example a background whose embedding shifts dramatically when the foreground is present, or where the optimal tau swings widely across contexts. Concretely, take a mixture, remove the measured background contribution at the paper's fixed tau, and check whether the refined profile's top-1 accuracy still tracks the SNR curve; if a single tau cannot serve both a masked and an unmasked background at the same SNR, the additivity premise fails.

Watch

Extended reading notes

Core claim

The central claim is that the influence of a background soundscape on a prototypical audio-text classifier is largely additive in the space of cosine-similarity profiles, so it can be removed by computing the background's own profile and subtracting it, scaled by a single parameter tau, from the test profile. With text-based background profiles the method consistently beats the non-adapted baseline, and with audio-based background profiles it beats it by a larger margin, in every tested environment and at 6, 8, and 10 dB SNR. The same subtraction works whether prototypes are text anchors or TGAP centroids estimated from unlabeled audio, and it transfers to real-world TUT data where audio adaptation lifts accuracy from 32.4% to 42.4%. The paper further claims that the modality gap, not just background interference, is a major limiter: prototypes that sit closer to audio embeddings yield systematically better classification, and the TGAP procedure narrows that gap without labels.

Load-bearing premise

The load-bearing premise is that a background soundscape's effect on the similarity profile is additive and can be cancelled by subtracting one fixed scaled background profile, regardless of how the foreground and background interact acoustically.

Editorial extensions

If this is right

  • Performance degradation from background soundscapes is driven primarily by SNR and is independent of background type, so a single adaptation recipe can serve many acoustic environments.
  • Subtracting a scaled background profile improves accuracy across all tested backgrounds and SNR levels without retraining the model.
  • Audio-based background profiles consistently outperform text-based profiles, so access to even a short background recording is more valuable than a textual description for domain adaptation.
  • Narrowing the audio-text modality gap, as TGAP does by building prototypes from unlabeled audio, improves classification on its own and compounds with background-profile subtraction.
  • The method transfers to real-world data: on TUT recordings, audio-based adaptation raises zero-shot accuracy from 32.4% to 42.4%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the additive-profile assumption holds broadly, the same subtraction could be applied iteratively to separate multiple overlapping sound sources, and the scalar tau could be predicted per recording instead of fixed globally.
  • In deployment without a clean background recording, one could estimate the background profile from the mixture itself or from a library of canonical scene profiles, which the paper does not explore.
  • The reported independence from background type suggests that sample-efficient adaptation is possible: a single background profile per SNR level could replace per-scene calibration.
  • A direct test of the additivity premise would be to measure whether the optimal tau remains stable when foreground and background are acoustically similar, since such cases stress the linear-subtraction assumption.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper studies zero-shot sound classification with audio-text models when the target recording contains a foreground sound mixed into a background soundscape. Using synthetic mixtures of UrbanSound8k foregrounds with TAU Urban Acoustic Scenes backgrounds, the authors show that accuracy drops with the presence of background and that the drop is mainly controlled by SNR rather than by the identity of the background. They propose a training-free domain adaptation method that computes a cosine-similarity 'profile' between the test audio and class prototypes, then subtracts a scaled background profile, obtained either from text prompts or from an actual background audio recording (Eq. 1: P_f = P_s - P_b * tau). The method is evaluated with text-anchor zero-shot prototypes and with TGAP audio prototypes, reporting consistent accuracy improvements across all six background types and SNR levels, with audio-based adaptation giving the largest gains; a TUT-based real-world check is also reported. In addition, the paper analyzes the modality gap, showing that TGAP centroids reduce the cosine distance between audio samples and class prototypes compared with text anchors.

Significance. If the reported effects are genuine and the evaluation is not oracle-based, the contribution is useful and practical: the method is simple, requires no retraining, works with both text and audio background information, and comes with released code. The empirical analysis of how background soundscapes affect prototype-based classification and the comparison of text anchors with TGAP prototypes are relevant to the growing literature on audio-text models. The paper also includes out-of-domain validation on TUT, which is a strength. However, the central quantitative claims currently rest on an evaluation protocol in which the adaptation parameter is selected on the test data and in which audio-based adaptation may have access to the exact interfering background recording; these issues need to be resolved before the claims can be accepted at face value.

major comments (4)
  1. [Section 5] The central claim that the adaptation method 'consistently improves performance' is not supported as a predictive claim because tau is selected on the same evaluation data. Section 5 states that tau was optimized by grid search (0 to 1 in 0.1 steps) on the same test conditions used to report Table 2 and Figure 4. The reported improvements are therefore post-hoc fits to the test set rather than results of a fully specified adaptation method. The authors should select tau on a validation split or fix it a priori, and should report sensitivity of the results to tau around the chosen values.
  2. [Sections 2 and 3] The audio-based adaptation may rely on oracle access to the exact background signal. Section 2 constructs each test soundscape by pairing a foreground with one random background sample, and Section 3 defines the audio-based background profile from 'an actual background audio recording', but the paper never states whether this recording is the same file as the one mixed into the test clip. If it is, the large gains in Table 2 (e.g., TGAP Park from 67.7 to 72.2) and the TUT results in Section 5 overstate what a deployable system can achieve, since in practice the interfering signal is not separately available. The same ambiguity applies to the TUT experiment, where chunks without class presence are used as background and may come from the same recording as the test chunks. The authors should clarify this and provide a non-oracle evaluation using a background recording from the same scene type that is not the exact file used in the test mixture.
  3. [Tables 2 and Figure 4] No statistical evaluation is provided. All accuracy figures are single point estimates with no confidence intervals, no multiple random pairings of foregrounds and backgrounds, and no significance tests. Given the small differences between Text and Audio adaptation in several rows of Table 2 (e.g., ZS Street Traffic: 66.8 vs 68.8), it is unclear whether the ordering of methods is reliable. Reporting variance across repeated synthetic dataset generations or at least per-background variability would materially strengthen the claims.
  4. [Equation (1)] The additive subtraction assumption in Eq. (1) is not directly tested. The method assumes that the influence of a background soundscape on the profile is a scaled additive component that can be removed with a single tau fixed across backgrounds and SNRs. The paper does not provide diagnostics such as the residual profile after subtraction compared with the foreground-only profile, nor does it analyze how the optimal tau varies per background and SNR. Since the authors concede that 'the optimal value showed slight variations depending on the specific context and SNR', a fixed-tau analysis with per-condition optima would clarify whether the chosen 0.2 and 0.7 are representative or merely the best average on the test set.
minor comments (5)
  1. [Section 5] The text refers to 'TUT datasets (DCASE SED Challenge 2016 [21] and 2017 [22])' but reports only a single TUT accuracy; please specify which dataset and split was used or whether results are pooled across both.
  2. [Section 2, Figure 1] The claim that degradation is 'independent of background type' is stronger than the visual evidence, since accuracy at the same SNR varies by several points across backgrounds (e.g., 61.0 for Park vs 55.8 for Shopping Mall at 6 dB). Please soften to 'primarily driven by SNR' or provide a statistical test of a background-type effect.
  3. [Section 2] Reference [19] is Radford et al., 'Robust speech recognition via large-scale weak supervision', which does not appear to support the statement that SNR values below 6 dB or above 10 dB have particular perceptual or classification properties; please supply a relevant reference or remove the citation.
  4. [Experimental setup] The paper does not state which LION-CLAP checkpoint is used or how audio and text embeddings are computed. Adding this reproducibility detail would help readers apply the method.
  5. [Equation (1)] The variables P_s, P_b, and P_f are described verbally but not formally defined; please state explicitly that they are vectors of cosine similarities and describe any normalization applied before the subtraction.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported gains rely on τ grid-searched on the same test set; audio-background profile source is underspecified.

  1. fitted input called prediction [Section 5, Results and discussion (τ grid search paragraph); Conclusions: 'An automatic selection of this parameter τ is left for future research.']
    "Regarding the parameter τ, we optimized it using a grid search (0 to 1 in 0.1 steps), identifying 0.2 for text-based and 0.7 for audio-based adaptation as generally optimal. ... The optimal value showed slight variations depending on the specific context and SNR’s."

    The accuracy gains in Table 2 and Figure 4 are produced after selecting τ by grid search on the same test soundscapes and SNR conditions. Thus the headline improvement is not a prediction of a fixed method but the result of fitting the subtraction weight to the evaluation data. The paper provides no validation split or principled derivation of τ, and explicitly defers automatic τ selection to future work, confirming that Eq. 1's weight is not independently determined. The reported increases (e.g., ZS Park 61.0 to 65.7; TGAP TUT 32.4 to 42.4 with audio adaptation) are therefore partly constructed by the grid search rather than being an independent demonstration.

full rationale

The central empirical claim (the method improves zero-shot prototypical classification under background soundscapes) is partially circular because the subtraction weight τ in Eq. 1 is optimized by grid search on the same evaluation conditions, so the Table 2 and Figure 4 numbers are in-sample fits. The paper is transparent about this and leaves automatic τ selection to future work, but this means the reported gains do not constitute a fixed-parameter prediction. The co-authored TGAP citation ([15], Kushwaha and Fuentes) is not treated as load-bearing circularity: TGAP is prior work, and the paper supplies independent cosine-distance evidence (Figure 3) for its modality-gap effect. The additional concern that audio-based background profiles might use the exact background recording mixed into each test clip cannot be established from the manuscript text alone, since the paper does not specify the recording source in the synthetic setup; that would be an evaluation-leakage issue rather than a demonstrated circular reduction. Because the main accuracy improvements are partly an artifact of fitting τ to the test data, a score of 6 is warranted.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method's central mechanism assumes that background influence is additive in the cosine-similarity profile space and removable with a single tuned scale tau; the audio-based variant may rely on background recordings from the same distribution as the test mixtures. No new physical entities are introduced.

free parameters (2)
  • tau (background subtraction strength) = 0.2 for text-based, 0.7 for audio-based
    Grid search over 0 to 1 in 0.1 steps on the same test soundscapes (Section 5); the reported improvements depend on this tuned value.
  • prompt averaging configuration for text-based background profile = average of three prompt templates
    The authors selected the prompt set that gave the best results ('best results were obtained by averaging the profiles of three prompts'), an ad hoc choice not validated on held-out data.
assumptions (3)
  • domain assumption The background's influence on the cosine-similarity profile is additive and can be removed by scaled subtraction (Eq. 1).
    No derivation or empirical validation of additivity is provided; the method is built on this linearity assumption.
  • domain assumption Cosine similarity between audio and text embeddings is a meaningful measure of semantic class proximity for classification.
    Standard assumption in prototype-based audio-text models, invoked throughout Sections 2 through 4.
  • domain assumption Synthetically mixed soundscapes from Scaper with UrbanSound8k foreground and TAU backgrounds at SNR 6-10 dB are representative of real-world acoustic scenes.
    Generalization to real-world settings is claimed from this synthetic setup (Section 2), with one additional TUT validation (Section 5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Domain Adaptation Method and Modality Gap Impact in Audio-Text Models for Prototypical Sound Classification." pith.science (2026). https://pith.science/paper/OBDBYYKE

@misc{pith2026250604376,
  author       = {Pith},
  title        = {Pith review of: Domain Adaptation Method and Modality Gap Impact in Audio-Text Models for Prototypical Sound Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OBDBYYKE}},
  note         = {Machine review of arXiv:2506.04376}
}
read the original abstract

Audio-text models are widely used in zero-shot environmental sound classification as they alleviate the need for annotated data. However, we show that their performance severely drops in the presence of background sound sources. Our analysis reveals that this degradation is primarily driven by SNR levels of background soundscapes, and independent of background type. To address this, we propose a novel method that quantifies and integrates the contribution of background sources into the classification process, improving performance without requiring model retraining. Our domain adaptation technique enhances accuracy across various backgrounds and SNR conditions. Moreover, we analyze the modality gap between audio and text embeddings, showing that narrowing this gap improves classification performance. The method generalizes effectively across state-of-the-art prototypical approaches, showcasing its scalability and robustness for diverse environments.

Figures

Figures reproduced from arXiv: 2506.04376 by the authors.

Figure 1
Figure 1. Impact of background soundscape on zero-shot sound classification accuracy across different acoustic environments. The plot compares model performance under 6 dB, 8 dB, and 10 dB SNR conditions against the reference performance (i.e. isolated sources). In addition to evaluating downstream task performance, we measured the cosine distance between text anchors and gener￾ated soundscapes, which averaged 0.76–slightly h… view at source ↗
Figure 2
Figure 2. Proposed method for domain adaptation in prototypical sound classification. The method quantifies the background’s contri￾bution to the test audio soundscape and refines class estimation by subtracting its influence. First, the test audio profile is calculated by measuring the similarity between the test audio embedding and the class prototypes. Next, a background profile is generated, either from text descriptions … view at source ↗
Figure 3
Figure 3. Comparison of cosine distances between audio sam￾ples and three types of class prototypes across multiple sound classes: text-based anchors (red), TGAP estimated centroids (blue), and real audio centroids (yellow) in UrbanSound8k. The experiment shows that text anchors prototypes (red) have the highest cosine distances, indicating poor alignment with audio samples. TGAP prototypes (blue) significantly re￾duce this d… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Average performance across all backgrounds under varying SNRs. The Zero-Shot method is shown on the left, and TGAP on the right, with Baseline, Text, and Audio evaluation setups in each. In brief, our proposed DA method proves effective across different background envi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 18 canonical work pages

  1. [1]

    the sound of a park

    Introduction Audio-text models (ATMs), i.e. embedding models with seman- tic knowledge of language and sound, have become popular in zero-shot (ZS) sound classification as they relieve the need for collecting and annotating audio recordings [1, 2]. Aligning au- dio and text modalities facilitates prototype approaches, where anchors, specific embedding poi...

    work page

  2. [2]

    background

    Impact of Background Sounds To assess the impact of background sound on ZS sound clas- sification using ATMs, we synthetically generated soundscapes using theScraperlibrary [16]. The foreground samples were drawn from UrbanSound8k [17], while the background sound- scape samples were taken from TAU Urban Acoustic Scenes 2019 [18]. We generated new soundsca...

    arXiv 2019

  3. [3]

    contribution

    Method Proposed Sound classification task in this context implies accurately iden- tifying a foreground sound embedded within a background soundscape (e.g., detecting a dog barking in a metro station soundscape). The audio to be classified consists of a mixture of the foreground sound and the background soundscape, which we refer to as test audio soundsca...

    work page

  4. [4]

    Apart from this, ATMs also face the challenge of themodality gap, wherein embeddings from differ- ent modalities are separated into distinct regions within the la- tent space

    On the Audio-Text Modality Gap So far, we have mentioned how prototypical approaches in ATMs struggle to represent polyphonic acoustic scenes and to effectively leverage text. Apart from this, ATMs also face the challenge of themodality gap, wherein embeddings from differ- ent modalities are separated into distinct regions within the la- tent space. Despi...

    work page

  5. [5]

    We compare three settings: no do- main adaptation (Baseline), text-based adaptation (Text), and audio-based adaptation (Audio)

    Results and discussion The results of our proposed method for ZS sound classifica- tion are shown in Table 2. We compare three settings: no do- main adaptation (Baseline), text-based adaptation (Text), and audio-based adaptation (Audio). Both adaptation approaches consistently improve performance across all background envi- ronments, with audio-based adap...

    work page 2016

  6. [6]

    To address these challenges, we propose a domain adaptation method that significantly improves classification accuracy in various background environments and SNR levels

    Conclusions and future work We show how prototypical audio-text models face challenges in coherently representing background as a concept, which leads to confusion and performance degradation when classifying sounds in real-world settings. To address these challenges, we propose a domain adaptation method that significantly improves classification accurac...

    work page

  7. [7]

    Audioclip: Extend- ing clip to image, text and audio,

    A. Guzhov, F. Raue, J. Hees, and A. Dengel, “Audioclip: Extend- ing clip to image, text and audio,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2022, pp. 976–980

    2022

  8. [8]

    Clap learning audio concepts from natural language supervision,

    B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang, “Clap learning audio concepts from natural language supervision,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

Show all 28 references
  1. [9]

    Analysis of sampling methodologies for noise pollution assessment and the impact on the population,

    G. Rey Gozalo and J. M. Barrig´on Morillas, “Analysis of sampling methodologies for noise pollution assessment and the impact on the population,”International journal of environmental research and public health, vol. 13, no. 5, p. 490, 2016

    2016

  2. [10]

    Audio surveillance of roads: A system for detecting anomalous sounds,

    P. Foggia, N. Petkov, A. Saggese, N. Strisciuglio, and M. Vento, “Audio surveillance of roads: A system for detecting anomalous sounds,”IEEE Transactions on Intelligent Transportation Sys- tems, vol. 17, no. 1, pp. 279–288, 2016

    2016

  3. [11]

    “Not There Yet

    J. Z. Huang, H. Chhabria, and D. Jain, ““Not There Yet”: Feasibility and Challenges of Mobile Sound Recognition to Support Deaf and Hard-of-Hearing People,” inProceedings of the 25th International ACM SIGACCESS Conference on Computers and Accessibility, ser. ASSETS ’23. New Yo...

    2023

  4. [12]

    Deep learning and gradient boosting for urban environmental noise monitoring in smart cities,

    J. Renaud, R. Karam, M. Salomon, and R. Couturier, “Deep learning and gradient boosting for urban environmental noise monitoring in smart cities,”Expert Systems with Applications, vol. 218, p. 119568, 2023. [Online]. Available: https://www. sciencedirect.com/science/article/pi...

    2023

  5. [13]

    Large-scale contrastive language-audio pretrain- ing with feature fusion and keyword-to-caption augmentation,

    Y . Wu*, K. Chen*, T. Zhang*, Y . Hui*, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language-audio pretrain- ing with feature fusion and keyword-to-caption augmentation,” in IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP, 2023

    2023

  6. [14]

    Evalu- ation of deep audio representations for semantic sound similarity,

    R. O. Araz, D. Bogdanov, P. Alonso-Jim´enez, and F. Font, “Evalu- ation of deep audio representations for semantic sound similarity,” IN PRESS, 2024

    2024

  7. [15]

    Domain adapta- tion for contrastive audio-language models,

    S. Deshmukh, R. Singh, and B. Raj, “Domain adapta- tion for contrastive audio-language models,”arXiv preprint arXiv:2402.09585, 2024

    2024 arXiv

  8. [16]

    Palm: Few- shot prompt learning for audio language models,

    A. Hanif, M. T. Agro, M. A. Qazi, and H. Aldarmaki, “Palm: Few- shot prompt learning for audio language models,”arXiv preprint arXiv:2409.19806, 2024

    2024 arXiv

  9. [17]

    Audio-text mod- els do not yet leverage natural language,

    H.-H. Wu, O. Nieto, J. P. Bello, and J. Salamon, “Audio-text mod- els do not yet leverage natural language,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  10. [18]

    Its not a modality gap: Characterizing and addressing the contrastive gap,

    A. Fahim, A. Murphy, and A. Fyshe, “Its not a modality gap: Characterizing and addressing the contrastive gap,”arXiv preprint arXiv:2405.18570, 2024

    2024 arXiv

  11. [19]

    Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning,

    V . W. Liang, Y . Zhang, Y . Kwon, S. Yeung, and J. Y . Zou, “Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning,” inAdvances in Neural Infor- mation Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A....

    2022

  12. [20]

    Weakly-supervised auto- mated audio captioning via text only training,

    T. Kouzelis and V . Katsouros, “Weakly-supervised auto- mated audio captioning via text only training,”arXiv preprint arXiv:2309.12242, 2023

    2023 arXiv

  13. [21]

    A multimodal prototypical approach for unsupervised sound classification,

    S. S. Kushwaha and M. Fuentes, “A multimodal prototypical approach for unsupervised sound classification,”arXiv preprint arXiv:2306.12300, 2023

    2023 arXiv

  14. [22]

    Scaper: A library for soundscape synthesis and augmentation,

    J. Salamon, D. MacConnell, M. Cartwright, P. Li, and J. P. Bello, “Scaper: A library for soundscape synthesis and augmentation,” in2017 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2017, pp. 344–348

    2017

  15. [23]

    A dataset and taxonomy for urban sound research,

    J. Salamon, C. Jacoby, and J. P. Bello, “A dataset and taxonomy for urban sound research,” inProceedings of the 22nd ACM inter- national conference on Multimedia, 2014, pp. 1041–1044

    2014

  16. [24]

    A multi-device dataset for urban acoustic scene classification,

    A. Mesaros, T. Heittola, and T. Virtanen, “A multi-device dataset for urban acoustic scene classification,” inProceedings of the Detection and Classification of Acoustic Scenes and Events 2018 Workshop (DCASE2018), November 2018, pp. 9–13. [Online]. Available: https://arxiv.or...

    2018 arXiv

  17. [25]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  18. [26]

    Multi-label zero- shot audio classification with temporal attention,

    D. Dogan, H. Xie, T. Heittola, and T. Virtanen, “Multi-label zero- shot audio classification with temporal attention,”arXiv preprint arXiv:2409.00408, 2024

    2024 arXiv

  19. [27]

    TUT database for acoustic scene classification and sound event detection,

    A. Mesaros, T. Heittola, and T. Virtanen, “TUT database for acoustic scene classification and sound event detection,” in24th European Signal Processing Conference 2016 (EUSIPCO 2016), Budapest, Hungary, 2016

    2016

  20. [28]

    DCASE 2017 challenge setup: Tasks, datasets and baseline system,

    A. Mesaros, T. Heittola, A. Diment, B. Elizalde, A. Shah, E. Vin- cent, B. Raj, and T. Virtanen, “DCASE 2017 challenge setup: Tasks, datasets and baseline system,” inProceedings of the Detec- tion and Classification of Acoustic Scenes and Events 2017 Work- shop (DCASE2017), No...

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.