Pith. sign in

REVIEW 3 major objections 4 minor 46 references

ABHINAYA -- A System for Speech Emotion Recognition In Naturalistic Conditions Challenge

T0 review · 3 major / 4 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 An ensemble of five speech, text, and speech-text models claims 44.02% macro-F1 on the Naturalistic SER challenge test set, surpassing all published results.

desk verdict A credible challenge-system report with a solid official result; the post-challenge SoTA claim, however, is not yet checkable without a comparison table and error bars. read the letter →

arxiv 2505.18217 v1 pith:MHBMQ2Y5 submitted 2025-05-23 cs.SD cs.AIeess.AS

Soumya Dutta , Smruthi Balaji , Varada R , Viveka Salinamakki , Sriram Ganapathy This is my paper · ORCID
classification cs.SDcs.AIeess.AS
keywords speechemotionrecognitionnaturalisticconditionsensemblelearninglargelanguagemodelsclassimbalancemajorityvotinglossfunctionsself-supervised
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 presents Abhinaya, a system for speech emotion recognition in naturalistic conditions built for the Interspeech 2025 challenge. It claims that an ensemble of five models—two speech-based, two text-based, and one joint speech-text—combined by majority voting reaches a macro-F1 of 44.02% on the challenge test set, surpassing the 32.93% baseline and the best published result. The authors show that fine-tuning self-supervised models and speech large language models, together with a fine-tuned text model, captures emotional cues across modalities, and that per-modality loss functions for class imbalance improve results. The system ranked 4th among 166 submissions at the official deadline despite one model being only partially trained, and after that model finished training it achieved what the paper calls state-of-the-art performance.

What carries the argument

The load-bearing mechanism is the five-member ensemble with majority voting, where SALMONN-13B acts as tiebreaker. Each fine-tuned member uses attentive statistics pooling on last-layer representations and a loss chosen by validation: weighted focal loss for the speech-only models, vector scaling loss for the text and speech-text models. The speech-text SALMONN-7B model appends the Whisper-generated transcript to the audio input before LoRA fine-tuning, and its completed training is what lifts the ensemble from 41.81% to 44.02% test macro-F1.

What would settle it

Re-evaluate the Abhinaya ensemble on the challenge test set using reference transcripts instead of Whisper output; a macro-F1 near the speech-only baseline would show the text components add no true emotion signal. Alternatively, any published result above 44.02% macro-F1 on the same test set would directly falsify the paper's SoTA claim.

Watch

Extended reading notes

Core claim

The central claim is that a majority-voting ensemble of heterogeneous model types—WavLM-Large fine-tuned with weighted focal loss, SALMONN-13B fine-tuned with the same loss, a zero-shot LLaMA-3.3-70B text classifier, LLaMA-3.1-8B fine-tuned with vector scaling loss, and SALMONN-7B fine-tuned on concatenated speech plus Whisper-transcribed text—achieves 44.02% macro-F1 on the challenge test set. This exceeds the challenge baseline of 32.93% and, the paper states, the best published result on this task. The paper also shows that the full ensemble outperforms the best individual model by 24.6% relative, that the speech-text model adds roughly 2% absolute over the other four, and that the fine-tuned 8B text model beats the zero-shot 70B model.

Load-bearing premise

The load-bearing premise is that Whisper-generated ASR transcripts, used in place of reference transcripts during training and testing, preserve enough emotional content for the text and speech-text models to contribute genuine signal; if transcription errors systematically strip emotion cues, the ensemble's gains over the speech-only models would shrink.

Editorial extensions

If this is right

  • If the results hold, the Abhinaya ensemble is the best published method on the Interspeech 2025 Naturalistic SER test set, about 34% relative better than the challenge baseline.
  • Ensembling across modalities beats any single model: the full five-model vote gains 24.6% relative over the strongest individual system on the test set.
  • Joint speech-text modeling adds value beyond speech alone: the speech-text SALMONN-7B reaches 35.43% validation macro-F1 versus 33.87% for the speech-only 7B variant.
  • Loss function choice is modality-dependent: weighted focal loss suits the speech models while vector scaling suits the text and speech-text models, and mixing them helps the ensemble.

Reading between the lines

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

  • A natural ablation the paper does not run: retraining the text models on reference transcripts for training and validation would isolate whether Whisper's ASR output suppresses emotional cues, since the test-time dependence on Whisper is untested.
  • The SoTA figure is benchmark-specific; the transferable insight is that diverse modalities plus tailored losses beat any single backbone, which should be testable on other imbalanced SER corpora.
  • The fine-tuned 8B model outperforming the zero-shot 70B model suggests task-specific fine-tuning matters more than scale for text-side emotion recognition on ASR transcripts, a hypothesis that could be checked with other model families.
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

3 major / 4 minor

Summary. This manuscript describes Abhinaya, an ensemble system for the Interspeech 2025 Speech Emotion Recognition in Naturalistic Conditions Challenge. The system combines two speech-only models (WavLM-Large fine-tuned with weighted focal loss and SALMONN-13B fine-tuned with weighted focal loss), two text-only models (zero-shot LLaMA-3.3-70B and fine-tuned LLaMA-3.1-8B with vector scaling loss), and one speech-text SALMONN-7B model trained with vector scaling loss, with majority voting over the five classifiers. On the official challenge test set the full ensemble scored 41.81% macro-F1 and ranked 4th among 166 submissions, while an organizer-run post-challenge evaluation after completing ST1 fine-tuning gave 44.02% macro-F1. The paper reports ablations on loss functions, ensemble compositions, class-wise performance, and zero-shot LLM choices, and claims state-of-the-art performance among published results.

Significance. The main strength of the paper is that its headline numbers are externally evaluated: the official leaderboard score and the post-challenge organizer-run score are not author-computed test-set numbers, which avoids selection-on-test bias. The paper also releases code, provides a clear ablation of loss functions (Table 2) and ensemble combinations (Table 4), and documents class-wise behavior (Table 3), including the rare-class failure modes of speech-only models. If the SoTA claim is substantiated by a proper comparison with other published systems and by uncertainty quantification, the work would be a useful reference for applying SSL/SLLM/LLM ensembles and class-imbalance losses to naturalistic SER. At present, however, the claim is not yet falsifiable from the manuscript alone.

major comments (3)
  1. [§4.3, Table 1] The central claim that the final system performance of 44.02% 'surpasses the best published result in the challenge' is not supported by any comparison in the manuscript. The only external numbers given are the challenge baseline (32.93%) and the authors' own official 4th-place score (41.81%); no table, citation, or protocol states what the best published result is, whether it comes from official leaderboard submissions or post-challenge updates, or how the 44.02% post-challenge evaluation relates to it. Please add a comparison table with the published test macro-F1 scores of competing systems (with citations) and explicitly define the comparison protocol, so the SoTA claim is verifiable.
  2. [§4.3, Table 1; §4.1] The 44.02% macro-F1 is a point estimate on a test set of only 3,200 utterances (Section 4.1), and the paper provides no confidence interval, bootstrap, or significance test against the next-best published score. Since macro-F1 is averaged over eight classes, the effective sample size per class is small, and a difference of a few points may be within sampling noise. Please report a bootstrap confidence interval for the ensemble score and, if possible, a paired significance test against the best competing system, or at least state the per-class test counts and the uncertainty explicitly.
  3. [§3.2] The text-only and speech-text models are trained and tested on Whisper-Large-v3 ASR transcripts even though reference transcripts were available for training and validation. The paper does not quantify how ASR errors affect emotion cues, which is the load-bearing assumption behind using these modalities in the ensemble. Please add an analysis comparing (at least on validation) models trained and validated on reference transcripts versus ASR transcripts, including emotion-class-wise agreement, or otherwise justify that the ASR noise is emotion-preserving. Without this, the contribution of the text modalities to the ensemble and to the SoTA claim is not fully established.
minor comments (4)
  1. [Table 4] The symbols '#' and '!' are not defined, and their use is initially counterintuitive (with '!' indicating inclusion). Add a legend or use clearer notation such as check and cross marks.
  2. [References] Reference [16] is listed as 'Under submission'; use the final citation if the challenge overview paper is now published.
  3. [Figure 2] The zero-shot LLM validation scores in Figure 2 are reported without the exact prompt template or any indication of prompt sensitivity; a sentence describing the prompt would improve reproducibility.
  4. [Table 1] The test macro-F1 values for T1, T2, and ST1 are marked as not reported; since these components are part of the submitted ensemble, a sentence explaining why their individual test scores could not be reported would avoid ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports an externally evaluated benchmark result, not a derivation that reduces to its own inputs.

full rationale

The paper makes no formal derivation claim. Its central assertion, that the Abhinaya ensemble reaches 44.02% macro-F1 and surpasses previously published challenge results (Section 4.3), is an empirical measurement from an organizer-run post-challenge evaluation on test data whose labels were withheld from the authors (Section 4.1). No fitted parameter is renamed as a prediction: the loss hyperparameters (gamma, tau), LoRA rank, learning rate, and epoch counts are selected on a balanced validation set, which is standard model selection rather than circular reasoning. The self-citations ([17], [24], [27]) appear only as background for prior SER and multimodal work and are not load-bearing for the ensemble result. The paper does not invoke any uniqueness theorem, and no equation in the paper is equivalent by construction to a reported outcome. The valid concern that the SoTA claim is weakened by the absence of a full comparison table, a confidence interval, and explicit citations to competing published scores is a falsifiability and reporting limitation, not a circularity. Accordingly, the circularity burden is minimal and the score is 0.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The system is built from established pretrained models and standard losses; the measurable free parameters are training hyperparameters selected on the balanced validation set. No invented entities are introduced. The main domain assumptions are the reliability of the emotion labels, the fidelity of Whisper transcripts, and the representativeness of the balanced validation set.

free parameters (8)
  • WFL gamma = 2
    Hyper-parameter in weighted focal loss (Eq. 2); used for speech-only models S1 and S2 and selected on validation data (Section 4.2).
  • VS loss gamma = 0.3
    Temperature-scale exponent in vector scaling loss (Eq. 3); used for T2 and ST1 and selected on validation data (Section 4.2).
  • VS loss tau = 1
    Bias scaling coefficient in Eq. 3; used for T2 and ST1 and selected on validation data (Section 4.2).
  • LoRA rank r = 8
    Rank of LoRA adapters for S2, T2, and ST1; chosen globally and not ablated (Section 4.2).
  • LoRA alpha = 32
    Scaling factor for LoRA adapters; chosen globally and not ablated (Section 4.2).
  • Training epochs = 20 (ST1 at deadline: 3)
    All fine-tuned models trained for 20 epochs with best validation checkpoint; ST1 had only 3 epochs at the official deadline (Sections 4.2 and 4.3).
  • Learning rate = 1e-5
    AdamW learning rate for all fine-tuned models (Section 4.2).
  • Audio max duration = 10 seconds
    Speech and speech-text models process audio with a maximum duration of 10 seconds (Section 4.2).
assumptions (4)
  • domain assumption MSP-PODCAST emotion annotations are reliable ground truth for the 8 primary categories, with 'other' and 'X' excluded.
    Section 4.1: training and test labels are taken as given, and all reported performance numbers are computed relative to these labels.
  • domain assumption Whisper-large-v3 ASR transcripts carry enough emotional content for text and speech-text models, and ASR errors do not systematically harm emotion classification.
    Section 3.2: reference transcripts are available for training, but the authors use Whisper transcripts for consistency with the unavailable test transcripts.
  • domain assumption The balanced validation set (326 utterances per class) is representative of the balanced test set for model selection and ensemble choice.
    Section 4.1: the balanced validation set is constructed by the authors, and checkpoints plus majority-voting combinations are chosen on it.
  • domain assumption Pre-trained WavLM, SALMONN, and LLaMA representations are suitable initialization for emotion classification after fine-tuning.
    Section 3: all speech and text systems are built by fine-tuning these pretrained models rather than training from scratch.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ABHINAYA -- A System for Speech Emotion Recognition In Naturalistic Conditions Challenge." pith.science (2026). https://pith.science/paper/MHBMQ2Y5

@misc{pith2026250518217,
  author       = {Pith},
  title        = {Pith review of: ABHINAYA -- A System for Speech Emotion Recognition In Naturalistic Conditions Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHBMQ2Y5}},
  note         = {Machine review of arXiv:2505.18217}
}
read the original abstract

Speech emotion recognition (SER) in naturalistic settings remains a challenge due to the intrinsic variability, diverse recording conditions, and class imbalance. As participants in the Interspeech Naturalistic SER Challenge which focused on these complexities, we present Abhinaya, a system integrating speech-based, text-based, and speech-text models. Our approach fine-tunes self-supervised and speech large language models (SLLM) for speech representations, leverages large language models (LLM) for textual context, and employs speech-text modeling with an SLLM to capture nuanced emotional cues. To combat class imbalance, we apply tailored loss functions and generate categorical decisions through majority voting. Despite one model not being fully trained, the Abhinaya system ranked 4th among 166 submissions. Upon completion of training, it achieved state-of-the-art performance among published results, demonstrating the effectiveness of our approach for SER in real-world conditions.

Figures

Figures reproduced from arXiv: 2505.18217 by the authors.

Figure 1
Figure 1. Schematic of the different components of the Abhinaya SER system. We use three types of models - speech-only (S1, S2), text-only (T1, T2) and speech-text (ST1). Only T1 is used in a zero-shot setting. The text used by the models are ASR transcripts generated by Whisper [20]. Emotion recognition using LLMs: A comprehensive eval￾uation by Zhang et al. [26] highlights the effectiveness of LLMs in text sentiment classif… view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 37 canonical work pages

  1. [1]

    Abhinaya

    Introduction Speech carries significant para-linguistic information like cues about speaker, language, ambient conditions, emotions, and physiological/psychological health states of the speaker. The efforts to extract paralinguistic attributes has greatly benefited from evaluation campaigns and global benchmarking, much like other fields of machine learni...

    work page 2009

  2. [2]

    ABHINAYA -- A System for Speech Emotion Recognition In Naturalistic Conditions Challenge

    Background Speech foundation models for SER: Speech foundation mod- els trained in a self-supervised learning fashion such as wav2vec2.0 [21], HuBERT [22] or WavLM [18] yield rich speech representations for various tasks, including speech emo- tion recognition (SER). Both wav2vec2.0 and HuBERT embed- dings have been used for SER [23, 24, 25]. The WavLM-la...

    work page Pith review arXiv 2025

  3. [3]

    neutral”) almost26 times that of the minority class (“fear

    Proposed SER System The schematic of the Abhinaya system2 is shown in Fig. 1. 3.1. Speech-only models 3.1.1. Wav-LM SSL (S1) The speech foundation model used in this sub-system is the WavLM-large model [18]. The convolutional feature extractor is frozen, and transformer layers are fine-tuned. Since WavLM generates frame-level representations, we apply att...

    work page

  4. [4]

    happy”, “angry

    Experiments 4.1. Dataset The MSP-PODCAST [14] dataset serves as the basis for this challenge. The training, validation and the test splits con- tain 84260, 31961 and 3200 speech files respectively. The dataset includes annotations for 8 primary emotion categories: “happy”, “angry”, “sad”, “neutral”, “surprise”, “fear”, “con- tempt” and “disgust”. Addition...

    work page

  5. [5]

    Our approach com- bines five models leveraging large language models (LLMs) and speech large language models (SLLMs)—two speech-based, two text-based, and one multimodal model

    Conclusion In this paper, we present Abhinaya, a system for speech emo- tion recognition (SER) developed as part of the Interspeech Nat- uralistic Speech Emotion 2025 challenge. Our approach com- bines five models leveraging large language models (LLMs) and speech large language models (SLLMs)—two speech-based, two text-based, and one multimodal model. To...

    work page 2025

  6. [6]

    Two decades of speaker recognition eval- uation at the national institute of standards and technology,

    C. S. Greenberg et al., “Two decades of speaker recognition eval- uation at the national institute of standards and technology,”Com- puter Speech & Language, vol. 60, p. 101032, 2020

    work page 2020

  7. [7]

    The 2022 NIST Language Recognition Evaluation,

    Y . Lee et al., “The 2022 NIST Language Recognition Evaluation,” in Interspeech 2023, 2023, pp. 1928–1932

    work page 2022

  8. [8]

    The Third DIHARD Diarization Challenge,

    N. Ryant et al., “The Third DIHARD Diarization Challenge,” in Interspeech 2021, 2021, pp. 3570–3574

    work page 2021

Show all 46 references
  1. [9]

    Summary of the DISPLACE challenge 2023- DIarization of SPeaker and LAnguage in Conversational Environ- ments,

    S. Baghel et al., “Summary of the DISPLACE challenge 2023- DIarization of SPeaker and LAnguage in Conversational Environ- ments,” Speech Communication, vol. 161, p. 103080, 2024

    2023

  2. [10]

    DiCOV A Challenge: Dataset, Task, and Base- line System for COVID-19 Diagnosis Using Acoustics,

    A. Muguli et al., “DiCOV A Challenge: Dataset, Task, and Base- line System for COVID-19 Diagnosis Using Acoustics,” in Inter- speech 2021, 2021, pp. 901–905

    2021

  3. [11]

    The second DICOV A challenge: Dataset and performance analysis for diagnosis of COVID-19 using acous- tics,

    N. K. Sharma et al., “The second DICOV A challenge: Dataset and performance analysis for diagnosis of COVID-19 using acous- tics,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 556–560

    2022

  4. [12]

    The VOiCES from a Distance Challenge 2019,

    M. K. Nandwana et al., “The VOiCES from a Distance Challenge 2019,” in Interspeech 2019, 2019, pp. 2438–2442

    2019

  5. [13]

    CHiME-6 Challenge: Tackling Multispeaker Speech Recognition for Unsegmented Recordings,

    S. Watanabe et al., “CHiME-6 Challenge: Tackling Multispeaker Speech Recognition for Unsegmented Recordings,” in Proc. CHiME 2020, 2020, pp. 1–7

    2020

  6. [14]

    The INTERSPEECH 2009 emotion challenge,

    B. Schuller, S. Steidl, and A. Batliner, “The INTERSPEECH 2009 emotion challenge,” in Interspeech 2009, 2009, pp. 312–315

    2009

  7. [15]

    Avec 2011–the first international audio/visual emotion challenge,

    B. Schuller et al., “Avec 2011–the first international audio/visual emotion challenge,” in Affective Computing and Intelligent Inter- action: Fourth International Conference, ACII 2011, Proceed- ings, Part II. Springer, 2011, pp. 415–424

    2011

  8. [16]

    MUSE 2020 challenge and workshop: Multi- modal sentiment analysis, emotion-target engagement and trust- worthiness detection in real-life media: Emotional car reviews in- the-wild,

    L. Stappen et al., “MUSE 2020 challenge and workshop: Multi- modal sentiment analysis, emotion-target engagement and trust- worthiness detection in real-life media: Emotional car reviews in- the-wild,” in Proceedings of the 1st International on Multimodal Sentiment Analysis in...

    2020

  9. [17]

    IEMOCAP: Interactive emotional dyadic motion capture database,

    C. Busso et al., “IEMOCAP: Interactive emotional dyadic motion capture database,” LREC, vol. 42, no. 4, pp. 335–359, 2008

    2008

  10. [18]

    MSP-IMPROV: An acted corpus of dyadic interactions to study emotion perception,

    ——, “MSP-IMPROV: An acted corpus of dyadic interactions to study emotion perception,” IEEE Transactions on Affective Com- puting, vol. 8, no. 1, pp. 67–80, 2016

    2016

  11. [19]

    Building naturalistic emotionally bal- anced speech corpus by retrieving emotional speech from existing podcast recordings,

    R. Lotfian and C. Busso, “Building naturalistic emotionally bal- anced speech corpus by retrieving emotional speech from existing podcast recordings,” IEEE Transactions on Affective Computing, vol. 10, no. 4, pp. 471–483, 2017

    2017

  12. [20]

    Odyssey 2024 - Speech Emotion Recog- nition Challenge: Dataset, Baseline Framework, and Results,

    L. Goncalves et al. , “Odyssey 2024 - Speech Emotion Recog- nition Challenge: Dataset, Baseline Framework, and Results,” in The Speaker and Language Recognition Workshop (Odyssey 2024), 2024, pp. 247–254

    2024

  13. [21]

    The Interspeech 2025 Challenge on Speech Emotion Recognition in Naturalistic Conditions,

    A. Reddy Naini et al. , “The Interspeech 2025 Challenge on Speech Emotion Recognition in Naturalistic Conditions,” in In- terspeech 2025, vol. Under submission, Rotterdam, The Nether- lands, August 2025

    2025

  14. [22]

    Leveraging content and acoustic representations for efficient speech emotion recognition,

    S. Dutta and S. Ganapathy, “Leveraging content and acoustic representations for efficient speech emotion recognition,” arXiv preprint arXiv:2409.05566, 2024

    2024 arXiv

  15. [23]

    WavLM: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen et al., “WavLM: Large-scale self-supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  16. [24]

    SALMONN: Towards Generic Hearing Abilities for Large Language Models,

    C. Tang et al., “SALMONN: Towards Generic Hearing Abilities for Large Language Models,” in The Twelfth International Con- ference on Learning Representations, 2024

    2024

  17. [25]

    Robust speech recognition via large-scale weak supervision,

    A. Radford et al., “Robust speech recognition via large-scale weak supervision,” in International conference on machine learning . PMLR, 2023, pp. 28 492–28 518

    2023

  18. [26]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski et al., “wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in Neural Informa- tion Processing Systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  19. [27]

    HuBERT: Self-supervised speech representa- tion learning by masked prediction of hidden units,

    W.-N. Hsu et al., “HuBERT: Self-supervised speech representa- tion learning by masked prediction of hidden units,” IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  20. [28]

    Emotion recognition from speech using wav2vec 2.0 embeddings,

    L. Pepino, P. Riera, and L. Ferrer, “Emotion recognition from speech using wav2vec 2.0 embeddings,” Interspeech, pp. 3400– 3404, 2021

    2021

  21. [29]

    HCAM–Hierarchical Cross Atten- tion Model for Multi-modal Emotion Recognition,

    S. Dutta and S. Ganapathy, “HCAM–Hierarchical Cross Atten- tion Model for Multi-modal Emotion Recognition,”arXiv preprint arXiv:2304.06910, 2023

    2023 arXiv

  22. [30]

    Cross-corpus speech emotion recognition with hubert self-supervised representation,

    M. Pastor et al., “Cross-corpus speech emotion recognition with hubert self-supervised representation,” in INTERSPEECH 2022, 2022, pp. 76–80

    2022

  23. [31]

    Sentiment Analysis in the Era of Large Lan- guage Models: A Reality Check,

    W. Zhang et al., “Sentiment Analysis in the Era of Large Lan- guage Models: A Reality Check,” in Findings of the Association for Computational Linguistics: NAACL 2024 , 2024, pp. 3881– 3906

    2024

  24. [32]

    LLM supervised Pre-training for Multimodal Emotion Recognition in Conversations,

    S. Dutta and S. Ganapathy, “LLM supervised Pre-training for Multimodal Emotion Recognition in Conversations,” arXiv preprint arXiv:2501.11468, 2025

    2025 arXiv

  25. [33]

    Qwen-audio: Advancing universal audio under- standing via unified large-scale audio-language models,

    Y . Chu et al., “Qwen-audio: Advancing universal audio under- standing via unified large-scale audio-language models,” arXiv preprint arXiv:2311.07919, 2023

    2023 arXiv

  26. [34]

    WavLLM: Towards robust and adaptive speech large language model,

    S. Hu et al., “WavLLM: Towards robust and adaptive speech large language model,” in Findings of the Association for Computa- tional Linguistics: EMNLP 2024. Miami, Florida, USA: Associ- ation for Computational Linguistics, Nov. 2024, pp. 4552–4572

    2024

  27. [35]

    Focal loss for dense object detection,

    T.-Y . Lin et al., “Focal loss for dense object detection,” in ICCV, 2017, pp. 2980–2988

    2017

  28. [36]

    Label-imbalanced and group-sensitive classifi- cation under overparameterization,

    G. R. Kini et al., “Label-imbalanced and group-sensitive classifi- cation under overparameterization,”Advances in Neural Informa- tion Processing Systems, vol. 34, pp. 18 970–18 983, 2021

    2021

  29. [37]

    Attentive statistics pooling for deep speaker embedding,

    K. Okabe, T. Koshinaka, and K. Shinoda, “Attentive statistics pooling for deep speaker embedding,” in Interspeech 2018, 2018, pp. 2252–2256

    2018

  30. [38]

    BEATs: Audio Pre-Training with Acoustic To- kenizers,

    S. Chen et al., “BEATs: Audio Pre-Training with Acoustic To- kenizers,” in International Conference on Machine Learning . PMLR, 2023, pp. 5178–5193

    2023

  31. [39]

    LLaMA: Open and efficient foundation lan- guage models,

    H. Touvron et al., “LLaMA: Open and efficient foundation lan- guage models,” arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  32. [40]

    BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,

    J. Li et al., “BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,” in Inter- national Conference on Machine Learning . PMLR, 2023, pp. 19 730–19 742

    2023

  33. [41]

    The LLAMA 3 herd of models,

    A. Dubey et al., “The LLAMA 3 herd of models,” arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  34. [42]

    Decoupled weight de- cay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight de- cay regularization,” in 7th International Conference on Learning Representations, ICLR, , 2019. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  35. [43]

    LoRA: Low-Rank Adaptation of Large Language Models,

    E. J. Hu et al., “LoRA: Low-Rank Adaptation of Large Language Models,” inThe Tenth International Conference on Learning Rep- resentations, ICLR, 2022

    2022

  36. [44]

    DeepSeek-r1: Incentivizing reasoning ca- pability in llms via reinforcement learning,

    D. Guo et al. , “DeepSeek-r1: Incentivizing reasoning ca- pability in llms via reinforcement learning,” arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  37. [45]

    Gemini: a family of highly capable multimodal models,

    G. Team et al., “Gemini: a family of highly capable multimodal models,” arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  38. [46]

    GPT-4 technical report,

    J. Achiam et al. , “GPT-4 technical report,” arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

Pith tools

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