Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Dual Information Speech Language Models for Emotional Conversations

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A frozen LLM can perceive both words and tone when speech is split into two streams by small adapters, trained alone on public datasets, and still match or beat larger speech-language models on emotional conversation tasks.

desk verdict A useful adapter recipe for frozen-backbone speech LLMs, but the headline 'disentanglement' claim is not actually demonstrated by the experiments. read the letter →

arxiv 2508.08095 v1 pith:4LKC2PVA submitted 2025-08-11 cs.CL cs.AIcs.SDeess.AS

Chun Wang , Chenyang Liu , Wenze Xu , Weihong Deng This is my paper
classification cs.CLcs.AIcs.SDeess.AS
keywords speech-languagemodelparalinguisticinformationemotionrecognitionadapterequivalencereplacementregularizationfrozenLLMconversationalAIweaksupervision
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

This paper tries to show that a speech-language model for emotional conversation can be built without training a speech encoder or an LLM: only two small adapters are learned, and the two adapters keep linguistic and paralinguistic information in separate streams. The motivation is that earlier adapters entangle the two kinds of information and then degrade context understanding when they collapse into task-specific vectors. If the claim holds, emotional conversational AI can be assembled from frozen components plus a cheap adapter layer, on public datasets, and still match or beat models that update the whole encoder and LLM. The paper reports competitive scores on emotion, pitch, speed, and energy classification, near-ASR-quality transcription, and higher judge-rated conversational quality than two leading speech LLMs.

What carries the argument

The load-bearing mechanism is the dual-adapter architecture with heterogeneous output shapes. A paralinguistic adapter (a compact transformer followed by adaptive pooling to exactly $n_a = 10$ fixed embeddings) captures utterance-level style and acts as a soft prompt; a linguistic adapter (concatenating every $k = 5$ adjacent speech frames and passing them through two linear layers) produces a time-varying sequence that directly replaces the token embeddings of the transcript. These are trained by Equivalence Replacement Regularization (ERR), which randomly swaps or removes the other stream and demands unchanged task output, plus multi-turn conversation data with variable context positions.

What would settle it

Take utterances that differ only in a localized prosodic event (e.g., the same sentence with a sarcastic accent on the final word versus flat delivery, matched on mean pitch and energy). If the SLM with only paralinguistic embeddings rates the two as the same or responds neutrally in both cases, then the fixed-length pooling has discarded the temporally local cue and the disentanglement claim fails for such inputs.

Watch

Extended reading notes

Core claim

The central claim is that a frozen LLM can perceive both what is said and how it is said when the two kinds of information arrive through structurally different adapters. The paralinguistic adapter pools the speech encoder's output to a fixed length ($n_a = 10$ embeddings), matching the assumption that paralinguistic style is roughly utterance-level, while the linguistic adapter compresses the speech embedding sequence into a time-varying compact sequence that replaces the text embeddings the LLM would normally see. The two adapters are trained with Equivalence Replacement Regularization: for a linguistic task, the linguistic embeddings are randomly paired with paralinguistic embeddings sour

Load-bearing premise

The load-bearing premise is that an utterance's paralinguistic content (emotion, pitch, energy, tempo) is mostly uniform across time, so adaptive pooling to exactly ten fixed embeddings preserves what the model needs; cues that live in a single moment would be smoothed away.

Editorial extensions

If this is right

  • SLMs for emotion can be built with only adapter training; the speech encoder and LLM stay frozen, so their original abilities are preserved.
  • A single training recipe covers linguistic (ASR), paralinguistic (gender, pitch, speed, energy, emotion), and dual-information tasks, so no separate speech instruction-tuning corpus is required.
  • Judged conversational quality (content and style scores, emotional generation scores) can be higher than models that update the encoder and LLM while training far fewer parameters.
  • The consistency-under-replacement property gives a direct probe: if removing the paralinguistic stream changes a linguistic task's output, the adapters are entangled.
  • Because the adapters are small, the same frozen LLM can be swapped (the paper builds SLM-Qwen and SLM-Llama) with only adapter retraining.

Reading between the lines

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

  • The weakest link to test is the fixed-length assumption: if paralinguistic cues are localized (a sigh, a sarcastic accent on one word), pooling them to ten embeddings should fail; a targeted dataset with local prosodic events would settle where the boundary lies.
  • The same two-stream design could be ported to other frozen backbones and modalities (video gestures, facial expression) with the same ERR-style replacement supervision; the paper does not test this.
  • A cheaper interpretation of the paper's result is that a frozen LLM needs only a soft-prompt channel for style, not a full multimodal retraining, which suggests that emotional conversation is an interface problem rather than a reasoning problem.
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 / 6 minor

Summary. The paper proposes an SLM architecture that extends frozen LLMs with two lightweight adapters: a paralinguistic adapter that produces fixed-length embeddings via adaptive pooling and a linguistic adapter that produces downsampled temporal embeddings. Only the two adapters are trained, on LibriSpeech, TextrolSpeech, MELD, and StyleTalk/DailyTalk, using a three-stage instruction-tuning procedure. The core novelty is Equivalence Replacement Regularization (ERR), which randomly pairs one adapter's embeddings with embeddings derived from text, speech, or nothing, plus positional randomness during multi-turn training. The paper claims this disentangles paralinguistic and linguistic information and prevents task-specific vectors, preserving contextual understanding. Experiments report strong results on speech-attribute classification and MELD emotion recognition, competitive ASR WER on LibriSpeech, and higher LLM-judge scores than Qwen2-Audio and Llama-Omni on emotional conversation quality.

Significance. If the central claims were established, the paper would make a useful practical contribution: a frozen speech encoder and frozen LLM plus two small adapters could yield competitive emotional-conversation SLMs with parameter and data efficiency. The architecture is simple, mostly well specified, and evaluated on public benchmarks. However, the headline claims of information disentanglement and avoidance of task-specific vectors are not directly measured; they are inferred from downstream performance. The conversational evaluation relies on LLM judges from the same model families as the backbones and lacks human evaluation or significance testing. These gaps currently limit the strength of the conclusions, although the empirical results are suggestive and the approach is sufficiently concrete that the missing evidence could be supplied.

major comments (4)
  1. [Sec. III-A, Eq. (2)-(5); Sec. III-C] The disentanglement claim is not demonstrated. Both adapters consume the same speech encoder output ES, and ERR randomizes only the complementary adapter's input, not the adapter's own input. Therefore the paralinguistic adapter can encode linguistic content from ES, and the linguistic adapter can encode paralinguistic content, while still being robust when used alone (Tabs. I, III). The paper itself concedes in Sec. III-C that 'the linguistic adapter may encode time-dependent paralinguistic nuances, such as intonation.' To support the causal claim, the authors should directly probe the information content of P-Embs and L-Embs, e.g., train probes to predict linguistic content from P-Embs and paralinguistic attributes from L-Embs, or perform swap/neutralization experiments. Without such evidence, the mechanism advanced in the abstract and conclusion is unsupported.
  2. [Sec. III-C; Sec. V-B, Tab. IV] The claim that the proposed randomness prevents task-specific vectors and preserves context understanding is not backed by any controlled experiment. No ablation removes positional randomness or ERR sampling, nor is there any analysis of embedding stability across prompts/contexts. Table IV shows superior conversational scores, but this could arise from multi-turn data, task proportions, or the adapter architecture alone. The authors should ablate the randomness components and, ideally, analyze whether embeddings degenerate toward task-specific directions (e.g., via probing or similarity analysis).
  3. [Sec. V-B, Tab. IV; Sec. IV-B] The emotional-conversation evaluation uses only LLM judges from the same model families as the backbones: Qwen2.5-72B judges both models including SLM-Qwen (Qwen2.5-7B), and Llama-3.1-70B judges both including SLM-Llama (Llama-3.1-8B). Calling these 'independent' is misleading because same-family judges can exhibit systematic style or output-format preferences. No human evaluation or statistical significance/variance is reported. Since the conversational-quality claim is central, the authors should provide human ratings or at least a neutral-family judge, plus confidence intervals or significance tests.
  4. [Sec. III-A, Eq. (4)] The fixed-length paralinguistic representation na=10 assumes paralinguistic information is largely uniform across an utterance. This is a strong assumption for temporally localized cues (e.g., a sarcastic pitch accent or a sigh at one word). Even if disentanglement were established, the pooling discards the location of such cues. The paper does not analyze how na affects performance or whether temporally localized paralinguistic events are preserved. At minimum, this limitation should be stated explicitly and ideally tested by varying na or comparing with a temporally sensitive variant.
minor comments (6)
  1. [Tab. IV caption] The caption says 'RESULTS OF SLM-LLAMA ON THE STYLETALK DATASET' but the table contains both SLM-Qwen and SLM-Llama rows. Rename to reflect the contents.
  2. [Sec. III-A, Eq. (5)] The downsampling operation is described verbally but the notation h_i^S is never defined in an equation. Define the concatenation and the boundary conditions clearly.
  3. [Sec. III-C] The reader cannot tell how 'paralinguistic embeddings derived from text' are computed. Specify the text encoder or embedding procedure (e.g., from the speech caption text via the LLM embedding layer or a separate text encoder).
  4. [Sec. IV-A] The custom TextrolSpeech split is only described as 'customized train-validation-test splits'; details are in the appendix. Given that official test set is small, describe the split procedure and ensure no speaker/sample overlap is introduced between train and test.
  5. [Sec. IV-C] Hyperparameters are reported, but there is no discussion of sensitivity to the adapter hidden size dh, downsampling rate k, or na. A sensitivity analysis would strengthen the efficiency claim.
  6. [General] The appendix is hosted on Google Drive rather than as part of the arXiv submission; this may hinder reproducibility. Consider including an anonymized appendix in the submission.

Circularity Check

1 steps flagged · score 2.0 of 10

One trained-in ERR ablation presented as evidence; benchmark results are externally grounded.

  1. fitted input called prediction [Section V-A (Paralinguistic and Linguistic Tasks), p. 4; training defined in Section III-C Stage 2]
    "Furthermore, the results in Tab. I and Tab. III demonstrate that the proposed ERR effectively guides SLMs to perceive paralinguistic and linguistic information from their respective embeddings as intended. This is evidenced by the observation that on paralinguistic tasks (see Tab. I), both SLMs perform robustly when only paralinguistic embeddings (P-Embs) are present."

    Stage 2 trains the paralinguistic adapter on paralinguistic tasks while randomly pairing with paralinguistic embeddings from speech caption text, speech, or none, with equal probabilities. Hence the 'P-Embs only' condition is a training-time input combination; robustness under it is a direct consequence of the loss, not an independent emergent test. The same applies to L-Embs only on linguistic tasks. The paper presents this trained-in robustness as evidence for ERR-guided disentanglement, so this specific evidence reduces to the training objective. It does not show that P-Embs lack linguistic content, since both adapters consume the same ES (Eq. 2). This is a partial, non-load-bearing circularity in the mechanism claim; the external benchmark results (MELD, LibriSpeech, TextrolSpeech) rem

full rationale

The paper's core empirical results are evaluated on external, public benchmarks: MELD for emotion recognition, LibriSpeech for ASR, and TextrolSpeech for paralinguistic attributes, all with held-out splits. These give independent grounding and are not circular. There is no load-bearing self-citation: the cited prior work ([2], [5], [7], [8], [9], etc.) is not by the present authors, and no uniqueness theorem is imported. The conversation evaluation uses Qwen2.5-72B/Llama-3.1-70B judges, which are the same model families as the backbones; this is a potential evaluation bias, but it is not a definitional circularity because the judges are separate, larger models applied to held-out StyleTalk data. The one genuine circularity concern is the ERR evidence: the paper trains the adapters to be robust to the removal of the complementary adapter's embeddings and then cites that robustness as evidence that the adapters 'perceive information from their respective embeddings as intended.' That is a trained-in property, not a free-standing result. The paper itself weakens the full disentanglement claim by acknowledging: 'the adopted ERR does not strictly train the two adapters independently, acknowledging the challenges of complete disentanglement while allowing flexibility. For instance, the linguistic adapter may encode time-dependent paralinguistic nuances, such as intonation.' This, together with the fact that both adapters receive the same speech encoder output ES, means the claimed disentanglement is under-supported; but this is a correctness/evidence concern rather than a derivation that reduces entirely to its inputs. The main contribution is still supported by independent benchmark numbers, so the overall circularity is low.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method is an empirical architecture study; it relies on a fixed pretrained speech encoder and LLM, and the central premise is the constant-within-utterance nature of paralinguistic info. The listed free parameters are hand-chosen hyperparameters, not fitted to the reported evaluation losses, so they do not by themselves make the claims circular, but they are not derived from any principle.

free parameters (4)
  • na (number of paralinguistic embeddings) = 10
    Hand-chosen fixed length for pooled paralinguistic embeddings (Sec. IV-C); controls capacity of the paralinguistic adapter and is load-bearing for the fixed-length representation.
  • k (linguistic downsampling rate) = 5
    Speech embedding compression factor following [8]; chosen by hand, affects the temporal resolution of linguistic embeddings.
  • dh (hidden dim of linguistic adapter) = 2048
    Hidden layer dimension following [8]; hand-chosen, affects adapter capacity.
  • ERR sampling probabilities (speech caption / speech / none) = equal 1/3 each
    The three replacement sources are sampled with equal probabilities (Sec. III-C); the exact distribution is a design choice and is not varied or justified.
assumptions (5)
  • domain assumption Whisper-large-v3 provides sufficient speech representations for both ASR and attribute classification when frozen.
    The model relies on a frozen pretrained speech encoder (Sec. IV-C); if the encoder loses paralinguistic detail, the adapters cannot recover it.
  • domain assumption Paralinguistic information is mostly consistent throughout an utterance, so fixed-length pooled embeddings are sufficient.
    Stated in Sec. I ('mostly remains consistent') and Sec. III-A; this motivates adaptive pooling to na=10 and is the weakest premise.
  • domain assumption The LLM's input embedding space is text-prioritized, so linguistic and paralinguistic signals must be injected through separate mechanisms.
    Claimed in Sec. I: 'this text-trained space naturally prioritizes linguistic content, neglecting paralinguistic aspects.' If false, a single adapter would suffice.
  • domain assumption Responses generated by the underlying frozen LLM are suitable targets for the style-aware alignment task.
    Sec. IV-A states targets for dual information tasks are 'responses generated by the underlying LLM.' This assumes the base LLM behavior is the desired behavior for emotional conversation.
  • ad hoc to paper Task-specific vectors are harmful and the introduced randomness prevents them.
    The paper asserts this without directly measuring task-specific vectors (Sec. III-C); it is an untested mechanism central to the claimed context preservation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual Information Speech Language Models for Emotional Conversations." pith.science (2026). https://pith.science/paper/4LKC2PVA

@misc{pith2026250808095,
  author       = {Pith},
  title        = {Pith review of: Dual Information Speech Language Models for Emotional Conversations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4LKC2PVA}},
  note         = {Machine review of arXiv:2508.08095}
}
read the original abstract

Conversational systems relying on text-based large language models (LLMs) often overlook paralinguistic cues, essential for understanding emotions and intentions. Speech-language models (SLMs), which use speech as input, are emerging as a promising solution. However, SLMs built by extending frozen LLMs struggle to capture paralinguistic information and exhibit reduced context understanding. We identify entangled information and improper training strategies as key issues. To address these issues, we propose two heterogeneous adapters and suggest a weakly supervised training strategy. Our approach disentangles paralinguistic and linguistic information, enabling SLMs to interpret speech through structured representations. It also preserves contextual understanding by avoiding the generation of task-specific vectors through controlled randomness. This approach trains only the adapters on common datasets, ensuring parameter and data efficiency. Experiments demonstrate competitive performance in emotional conversation tasks, showcasing the model's ability to effectively integrate both paralinguistic and linguistic information within contextual settings.

Figures

Figures reproduced from arXiv: 2508.08095 by the authors.

Figure 1
Figure 1. An illustrative example of how understanding paralinguistic informa [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the SLM model architecture. Learnable modules in red, [PITH_FULL_IMAGE:figures/full_fig_p001_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the Equivalence Replacement Regularization. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 20 canonical work pages

  1. [1]

    Towards emotional support dialog systems,

    Siyang Liu et al., “Towards emotional support dialog systems,” in Proceedings of the 59th Annual Meeting of the Association for Com- putational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Aug. 2021, pp. 3469–3483

    work page 2021

  2. [2]

    Advancing large language models to capture varied speaking styles and respond properly in spoken conversations,

    Guan-Ting Lin, Cheng-Han Chiang, and Hung-yi Lee, “Advancing large language models to capture varied speaking styles and respond properly in spoken conversations,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Bangkok, Thailand, Aug. 2024, pp. 6626–6642

    work page 2024

  3. [3]

    Paralinguistics-enhanced large language modeling of spoken dialogue,

    Guan-Ting Lin et al., “Paralinguistics-enhanced large language modeling of spoken dialogue,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 10316–10320

    work page 2024

  4. [4]

    E-chat: Emotion-sensitive spoken dialogue system with large language models,

    Hongfei Xue et al., “E-chat: Emotion-sensitive spoken dialogue system with large language models,” arXiv:2401.00475, 2024

    arXiv 2024

  5. [5]

    Blsp-emo: Towards empathetic large speech- language models,

    Chen Wang et al., “Blsp-emo: Towards empathetic large speech- language models,” arXiv:2406.03872, 2024

    arXiv 2024

  6. [6]

    Wavchat: A survey of spoken dialogue models,

    Shengpeng Ji et al., “Wavchat: A survey of spoken dialogue models,” arXiv:2411.13577, 2024

    arXiv 2024

  7. [7]

    Connecting speech encoder and large language model for asr,

    Wenyi Yu et al., “Connecting speech encoder and large language model for asr,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2024, pp. 12637– 12641

    work page 2024

  8. [8]

    An embarrassingly simple approach for llm with strong asr capacity,

    Ziyang Ma et al., “An embarrassingly simple approach for llm with strong asr capacity,” arXiv:2402.08846, 2024

    arXiv 2024

Show all 31 references
  1. [9]

    Frozen large language models can perceive paralinguistic aspects of speech,

    Wonjune Kang et al., “Frozen large language models can perceive paralinguistic aspects of speech,” arXiv:2410.01162, 2024

    2024 arXiv

  2. [10]

    Prefix-tuning: Optimizing continuous prompts for generation,

    Xiang Lisa Li and Percy Liang, “Prefix-tuning: Optimizing continuous prompts for generation,” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long P...

    2021

  3. [11]

    AudioChatLlama: Towards general-purpose speech abilities for LLMs,

    Yassir Fathullah et al., “AudioChatLlama: Towards general-purpose speech abilities for LLMs,” in Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , Mexico City...

    2024

  4. [12]

    P-tuning v2: Prompt tuning can be comparable to fine- tuning universally across scales and tasks,

    Xiao Liu et al., “P-tuning v2: Prompt tuning can be comparable to fine- tuning universally across scales and tasks,” arXiv:2110.07602, 2022

    2022 arXiv

  5. [13]

    Secap: Speech emotion captioning with large language model,

    Yaoxun Xu et al., “Secap: Speech emotion captioning with large language model,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 17, pp. 19323–19331, Mar. 2024

    2024

  6. [14]

    Stylecap: Automatic speaking-style captioning from speech based on speech and language self-supervised learning models,

    Kazuki Yamauchi, Yusuke Ijima, and Yuki Saito, “Stylecap: Automatic speaking-style captioning from speech based on speech and language self-supervised learning models,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 202...

    2024

  7. [15]

    Qwen2-audio technical report,

    Yunfei Chu et al., “Qwen2-audio technical report,” arXiv:2407.10759, 2024

    2024 arXiv

  8. [16]

    The power of scale for parameter-efficient prompt tuning,

    Brian Lester et al., “The power of scale for parameter-efficient prompt tuning,” in Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , Nov. 2021, pp. 3045–3059

    2021

  9. [17]

    Multitask prompt tuning enables parameter-efficient transfer learning,

    Zhen Wang et al., “Multitask prompt tuning enables parameter-efficient transfer learning,” in The Eleventh International Conference on Learning Representations, 2023

    2023

  10. [18]

    Factor-conditioned speaking-style captioning,

    Atsushi Ando, Takafumi Moriya, Shota Horiguchi, and Ryo Masumura, “Factor-conditioned speaking-style captioning,” in Interspeech 2024 , 2024, pp. 782–786

    2024

  11. [19]

    Developing instruction-following speech language model without speech instruction-tuning data,

    Ke-Han Lu, , et al., “Developing instruction-following speech language model without speech instruction-tuning data,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  12. [20]

    Textrolspeech: A text style control speech corpus with codec language text-to-speech models,

    Shengpeng Ji et al., “Textrolspeech: A text style control speech corpus with codec language text-to-speech models,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 10301–10305

    2024

  13. [21]

    MELD: A multimodal multi-party dataset for emotion recognition in conversations,

    Soujanya Poria et al., “MELD: A multimodal multi-party dataset for emotion recognition in conversations,” inProceedings of the 57th Annual Meeting of the Association for Computational Linguistics , Florence, Italy, July 2019, pp. 527–536

    2019

  14. [22]

    Dailytalk: Spoken dialogue dataset for conversational text-to-speech,

    Keon Lee, Kyumin Park, and Daeyoung Kim, “Dailytalk: Spoken dialogue dataset for conversational text-to-speech,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  15. [23]

    DailyDialog: A manually labelled multi-turn dialogue dataset,

    Yanran Li et al., “DailyDialog: A manually labelled multi-turn dialogue dataset,” in Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Taipei, Taiwan, Nov. 2017, pp. 986–995

    2017

  16. [24]

    Llama-omni: Seamless speech interaction with large language models,

    Qingkai Fang et al., “Llama-omni: Seamless speech interaction with large language models,” arXiv:2409.06666, 2024

    2024 arXiv

  17. [25]

    Enhancing emotional generation capability of large language models via emotional chain-of-thought,

    Zaijing Li et al., “Enhancing emotional generation capability of large language models via emotional chain-of-thought,” arXiv:2401.06836, 2024

    2024 arXiv

  18. [26]

    Robust speech recognition via large-scale weak supervision,

    Alec Radford et al., “Robust speech recognition via large-scale weak supervision,” in Proceedings of the 40th International Conference on Machine Learning, 23–29 Jul 2023, vol. 202 of Proceedings of Machine Learning Research, pp. 28492–28518

    2023

  19. [27]

    Qwen2.5: A party of foundation models,

    Qwen Team, “Qwen2.5: A party of foundation models,” September 2024

    2024

  20. [28]

    The llama 3 herd of models,

    Abhimanyu Dubey et al., “The llama 3 herd of models,” arXiv:2407.21783, 2024

    2024 arXiv

  21. [29]

    emotion2vec: Self-supervised pre-training for speech emotion representation,

    Ziyang Ma et al., “emotion2vec: Self-supervised pre-training for speech emotion representation,” in Findings of the Association for Computational Linguistics: ACL 2024 , Bangkok, Thailand, Aug. 2024, pp. 15747–15760

    2024

  22. [30]

    Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,

    Yunfei Chu et al., “Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,” arXiv:2311.07919, 2023

    2023 arXiv

  23. [31]

    SALMONN: Towards generic hearing abilities for large language models,

    Changli Tang et al., “SALMONN: Towards generic hearing abilities for large language models,” in The Twelfth International Conference on Learning Representations, 2024

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.