EmpaAva: An Open-source Agentic 3D-Avatar Empathetic Live Chatbot

Hao Fei; Jie Yang; Shuhui Lin; Wenhao Xu

REVIEW 4 major objections 4 minor 26 references

EmpaAva: An Open-source Agentic 3D-Avatar Empathetic Live Chatbot

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read EmpaAva is presented as the first open-source live 3D avatar chatbot that senses user emotion and answers with an empathetic, embodied reply.

desk verdict A genuinely novel open-source 3D-avatar empathetic chatbot whose central 'surpasses' claim currently rests on an ambiguous evaluation protocol that must be clarified before the numbers can be trusted. read the letter →

arxiv 2608.04709 v1 pith:BGFKTFYD submitted 2026-08-05 cs.CL

Jie Yang , Wenhao Xu , Shuhui Lin , Hao Fei This is my paper

classification cs.CL

keywords empatheticresponsegeneration3Davatartri-agentarchitectureplanningmultimodaldialogueLLMagentsspeechemotionrecognitionGaussianSplatting

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

EmpaAva is presented as the first open-source, live, agentic 3D-avatar empathetic chatbot: a video-call interface in which a user speaks to a 3D digital human that perceives the user's affect from speech and optional vision, plans an empathetic reply, and delivers it as emotional speech with lip-synced facial motion and photorealistic 3D Gaussian rendering. The paper's central claim is that a large language model coordinating a Tri-Agent Architecture — perception, response planning, embodied rendering — is what makes this loop work, with a Response Planning layer turning each reply into an executable multimodal plan that keeps wording, voice, and expression on one empathetic intent. Automatic and human evaluations reported in the paper place EmpaAva above text-only, 2D talking-face, and multimodal avatar baselines on emotion understanding, response quality, and audio-visual consistency. The authors release the implementation, prompts, an online demo, and the evaluation setup so the community can run and extend the system. The paper also states its own limits: speech emotion recognition can misread affect, and generated comfort can be generic.

What carries the argument

The central mechanism is the Tri-Agent Architecture plus the Response Planning layer. The architecture splits the long pipeline into PerceptionAgent, ResponseAgent, and RenderAgent, communicating through a shared human-readable dialogue state rather than opaque tensors, which makes each stage independently testable and swappable. The Response Planning layer is the load-bearing novelty: ResponseAgent outputs a structured reply plan (a JSON contract) that fixes who speaks, in which voice and emotional tone, against which background, and which modules to invoke, so every downstream generator obeys one expressive intent. The plan also cites the perception evidence it responds to, keeping the reply grounded and interpretable across turns.

What would settle it

Run EmpaAva's ResponseAgent on the AvaMERG test set with no emotion, cause, or reference text in its prompt, feeding only raw transcript, speech, and optional vision, then compare Emo.Acc and Cause.M to the reported 53.85 and 74.81; if accuracy drops to chance, the reported gains came from leaked labels rather than perception, and checking the released repository for the full live pipeline would settle whether the 'first open-source' claim is true.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that empathetic response generation can be carried out of the text channel and into a live, face-to-face 3D interaction by decomposing the perceive-decide-express chain into three cooperating agents coordinated by an LLM. PerceptionAgent fuses transcript, acoustic emotion, and optional visual frames into a user-state summary; ResponseAgent reasons over that state and emits a structured reply plan naming the reply text, target emotion and tone, avatar, voice, background, and the perception evidence each choice answers; RenderAgent executes the plan with emotional TTS, audio-driven FLAME motion, and 3D Gaussian Splatting. The paper reports that on the AvaMERG test set EmpaAva achieves the best diversity, emotion-understanding accuracy, and emotion-cause modeling among compared systems, and that human raters prefer its replies on empathy, relevance, and specificity. The intended contribution is the intelligence that binds existing open-source modules into one controllable, inspectable empathetic loop.

Load-bearing premise

The evaluation results depend on EmpaAva's response planner never being shown the dataset's true emotion labels or reference replies, so that its emotion-understanding and cause-modeling scores reflect genuine perception rather than leaked answers.

Editorial extensions

If this is right

Empathetic dialogue can be delivered as a live, embodied experience rather than text alone, because the reply plan couples language, voice, and facial expression to one intent.
Cross-modal consistency becomes a planning problem: if the LLM picks the emotion and tone first, TTS, lip motion, and rendering all inherit the same target.
Because the agents share text-form state, each stage can be upgraded independently — a better speech recognizer, a larger LLM, or a higher-fidelity renderer drops in without retraining the loop.
The open-source release with agent prompts gives the community a reproducible testbed for embodied empathetic agents.

Reading between the lines

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

If the reply-plan contract generalizes, the same structured intent could be extended to gesture, gaze, and proxemics, not just voice and face, since those channels are also downstream effectors of one planned emotion.
The reported end-to-end gains assume the automatic evaluation is a clean head-to-head; a direct test would withhold the AvaMERG ground-truth emotion labels from ResponseAgent's prompt and verify that Emo.Acc still comes from the perceived speech and vision.
The latency breakdown suggests live real-time use is not yet achieved: 45.8 seconds per turn dominated by Gaussian rendering and export, so reducing export time is the concrete lever before this becomes a true video-call substitute.
Because the paper states speech emotion recognition can misread affect and comfort can be generic, the near-term ceiling of the approach may be set by perception and content generation rather than by the 3D embodiment.

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.

Desk Editor's Note

read the letter

Real talk: this is a genuine system contribution with a real novelty, but the evaluation doesn't yet rule out two circularity risks that the abstract's 'surpasses' claim rides on.

What's actually new: the Tri-Agent Architecture and the Response Planning layer. The paper isn't claiming new TTS or new rendering; it's claiming an LLM-coordinated closed loop where perception, empathetic planning, and embodied rendering share one structured intent. That's a fair claim, and the system is open-source with a demo, which makes it a potentially useful testbed. The modularity and inspectability are real, and the reply-plan schema is a concrete, sensible artifact.

The soft spots are in the evaluation. First, the ResponseAgent backbone ambiguity. Appendix A, Table 5 lists the response stage as 'OpenAI-compatible LLM / AvaMERG fallback'. The paper never states which one generated the Figure 4 and Table 6 numbers. If the fallback was active during evaluation, then for those trials EmpaAva's responses are the exact model it's being compared against as the AvaMERG baseline, making the Emo.Acc gain (53.85 vs 30.97) circular. The authors need to say plainly that the fallback was disabled during evaluation, or rerun with the primary LLM.

Second, label leakage. Emo.Acc and Cause.M compare predicted labels to gold annotations in AvaMERG. The prompt templates in Appendix B don't show gold labels being inserted, but the manuscript nowhere states they were withheld from the perception summary or the LLM context. That's a one-line clarification, but it's load-bearing.

The human study is secondary but weak: 10 raters, 50 preference votes, no significance tests, and the gaps over AvaMERG are small (Empathy 4.42 vs 4.18, Preference 18/50 vs 13/50). It can't carry the 'surpasses' claim on its own.

Credit where due: the automatic results on EmpatheticDialogues (Table 2) are less exposed to the AvaMERG circularity risks, and the case studies are coherent. The system is actually built and runs, which is more than many papers at this level.

Who this is for: anyone building embodied empathetic agents or multimodal avatar systems. They'll get a clear reference architecture and a concrete planning layer to build on. But the central evaluation needs a clean retest before the numbers are cited.

Recommendation: send to peer review, not desk reject. A serious referee should demand (1) a statement of which response backbone produced each evaluation table, with the fallback disabled, (2) an explicit statement that gold labels were not available to the model during evaluation, and (3) ideally a larger, significance-tested human study. I'd read a revised version with those fixes.

Referee Report

4 major / 4 minor

Summary. The paper describes EmpaAva, an open-source, agentic 3D-avatar empathetic live chatbot that combines speech recognition, speech emotion recognition, an LLM-based Tri-Agent architecture (PerceptionAgent, ResponseAgent, RenderAgent), emotional TTS, audio-driven FLAME motion, and 3D Gaussian Splatting rendering. The central claim is that EmpaAva surpasses text-only, 2D talking-face, and multimodal avatar baselines in emotion understanding, response quality, and audio-visual consistency. Evidence includes automatic results on EmpatheticDialogues and the AvaMERG test set (Tables 2, 4, 6; Figure 4), a small human questionnaire (Table 3), and two multi-turn case studies. The system and an online demo are open-sourced.

Significance. If the evaluation claims hold, EmpaAva is a useful integration contribution: it demonstrates a complete, browser-deployable loop from affect perception to embodied 3D response, with a structured reply plan that makes cross-modal intent inspectable. The open-source release, detailed prompt templates in Appendix B, and coherent case studies are genuine strengths, and the automatic metrics consistently favor EmpaAva on most dimensions. However, the central 'surpasses baselines' claim depends on evaluation hygiene that is not yet documented: the exact response backbone is unspecified, gold-label leakage is not ruled out, and the human evaluation lacks statistical support. The significance is therefore conditional on resolving these points.

major comments (4)

[§5.1.2 / Appendix A, Table 5] The automatic end-to-end comparison (Figure 4, Table 6) is the load-bearing evidence for the claim that EmpaAva surpasses multimodal avatars in emotion understanding, but the paper never states which ResponseAgent backbone produced these numbers. Table 5 lists the Response stage as an 'OpenAI-compatible LLM / AvaMERG fallback,' and no experiment report indicates whether the fallback was active during evaluation. If the fallback is a variant of the AvaMERG baseline, then for those trials the evaluated responses would be generated by the comparison system itself, making the Emo.Acc (Eq. 1) and Cause.M (Eq. 3) gains partly circular. Please disclose the exact backbone and configuration for every evaluation run and rerun the comparison with the fallback disabled or clearly isolated.
[§5.1.2 / Appendix E / Appendix B] A second route to circularity is label leakage. Eqs. (1) and (3) compare predicted emotion and cause labels against the AvaMERG ground truth, but the manuscript nowhere states that the ground-truth emotion and cause annotations were withheld from the PerceptionAgent summary and the ResponseAgent prompt during evaluation. The prompt templates in Appendix B do not obviously include gold labels, but the absence of an explicit statement leaves open the possibility that the perceived emotion understanding partly encodes the test-set answer. Please state explicitly that gold labels were not available to any agent during evaluation, and ideally rerun the evaluation with a prompt that provably contains no gold-label text.
[§5.1.2 / Table 3] The human evaluation uses 10 raters and 50 preference votes, with no significance tests or confidence intervals. The reported gains are small (Empathy 4.42 vs 4.36; Preference 18/50 vs 13/50), and the text says the results 'echo the trend' rather than demonstrating it. Since response quality is a central part of the 'surpasses' claim, please report significance tests or confidence intervals, or soften the claim to a directional observation.
[§5.1.2 / §5.2 / Appendix E] The cross-modal consistency diagnostics MECS and 'A V-Agree' are reported only for EmpaAva and only on 100 outputs (22.0% and 40.0%), with no baseline values and no definition of 'A V-Agree' in Appendix E. As these diagnostics are used to support the audio-visual consistency claim, please provide baseline numbers, define the metric, or remove the implicit superiority conclusion drawn from these two numbers.

minor comments (4)

[Appendix E] The metric 'A-V Cos' and the diagnostic 'A V-Agree' are used in Tables 4 and §5.1.2 but are not defined or referenced; please add equations or citations for both.
[Figure 4] The figure caption says 'Higher scores indicate better response diversity, emotion understanding, and emotion-cause modeling,' but the axes and bar labels are not self-contained in the text; please ensure the figure is legible and clearly labeled.
[§5.2 / Table 4] The text states Wav2Lip wins LSE-D and EmpaAva gives the 'best overall sync–affect balance,' but no joint metric or significance test is provided; please report a combined index or clarify that this is a qualitative judgment.
[§5.1.1 / Table 2] The absolute Acc values on EmpatheticDialogues are low (best 10.26), which is not discussed; a brief note on why emotion-cue accuracy is in this range would help readers interpret the comparison.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation reduces to its own inputs; the central evaluation rests on external benchmarks, though shared authorship of two baselines and an unspecified AvaMERG fallback warrant transparency notes.

full rationale

Walking the paper's derivation chain, the central claim that EmpaAva surpasses text-only, 2D talking-face, and multimodal avatar baselines is supported by automatic and human evaluations on EmpatheticDialogues (Rashkin et al., 2019) and the AvaMERG benchmark (Zhang et al., 2025). EmpatheticDialogues is an external, widely used dataset, and AvaMERG, despite sharing an author with the present paper, is a separately published benchmark with its own released data. The automatic metrics in Eqs. (1) and (3) compare predicted emotion and cause labels to gold labels, and the prompt templates in Appendix B show the PerceptionAgent and ResponseAgent receive only transcript, speech-emotion (SER), optional visual cues, and dialogue history; no prompt template injects the gold emotion or cause labels, so no label-leakage path is specified or quoted. Appendix A, Table 5 lists the Response-stage backbone as 'OpenAI-compatible LLM / AvaMERG fallback,' but Table 6 reports EmpaAva scores (Emo.Acc 53.85, Cause.M 74.81) that differ substantially from the AvaMERG baseline (30.97 and 69.13), indicating the fallback was not active during the reported evaluation. The shared authorship of the AvaMERG and EmpathyEar baselines is a self-citation concern, but these are comparison systems, not premises from which EmpaAva's results are derived, so they are not load-bearing in a circularity sense. No equation is shown to equal its own input by construction, and no fitted parameter is renamed as a prediction. The unspecified fallback and the absence of an explicit statement that gold labels were withheld are transparency/correctness risks, not demonstrated circularity, so the score is a low 2 for minor self-citation rather than a higher circularity score.

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

The central claim rests on several domain assumptions about the reliability of speech emotion recognition, the ability of an LLM to emit executable reply plans, and the ability of the FLAME-to-Gaussian transfer to preserve identity. There are no physically invented entities. Two groups of hand-tuned rendering parameters affect the reported sync metrics but are not specified.

free parameters (2)

Expression/jaw/lip range clamps = not reported
Hand-set in Section 4.3 to keep motion expressive without tipping into exaggeration; the chosen values directly affect the audio-visual consistency scores in Table 4 but are not specified.
Temporal smoothing and offset correction settings = not reported
Applied in Section 4.3 to tighten lip-sync and suppress jitter; these settings influence LSE-C and LSE-D and are not quantified.

assumptions (3)

domain assumption Speech emotion recognition (emotion2vec/FunASR) provides a sufficiently reliable affect signal to condition empathetic response planning.
Invoked in Section 4.1; the paper itself notes in Ethical Considerations that SER can misread affect, so the empathetic loop's accuracy is bounded by this assumption.
domain assumption An LLM instructed with a JSON reply-plan schema will produce plans whose tone, expression, and voice fields keep the downstream generators aligned.
Core of Section 4.2; if the LLM emits inconsistent or ungrounded plans, the claimed cross-modal consistency fails.
domain assumption FLAME-to-Gaussian transfer can apply DEEPTalk-predicted motion to a rigged 3DGS head without breaking identity or visual quality.
Assumed in Section 4.3; the paper does not ablate this transfer step, so the rendering quality claim depends on it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EmpaAva: An Open-source Agentic 3D-Avatar Empathetic Live Chatbot." pith.science (2026). https://pith.science/paper/BGFKTFYD

@misc{pith2026260804709,
  author       = {Pith},
  title        = {Pith review of: EmpaAva: An Open-source Agentic 3D-Avatar Empathetic Live Chatbot},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BGFKTFYD}},
  note         = {Machine review of arXiv:2608.04709}
}

read the original abstract

This paper presents EmpaAva, to our knowledge the first open-source, agentic 3D-avatar empathetic chatbot, which carries empathetic response generation (ERG) from text-only exchanges into live, face-to-face interaction. Through a video-call-like interface, a user speaks to a 3D digital human that reads their affect from speech and optional vision, and replies with emotional speech, lip-synced facial motion, and photorealistic 3D Gaussian rendering. At its core, an LLM coordinates a Tri-Agent Architecture, in which perception, empathetic response planning, and embodied rendering form a closed loop, paired with a Response Planning layer that compiles each reply into an executable multimodal plan, keeping voice, expression, and rendering on one empathetic intent. Building on strong open-source modules, EmpaAva supplies the intelligence that binds them into one controllable, inspectable experience. In automatic and human evaluations, EmpaAva surpasses text-only, 2D talking-face, and multimodal avatar baselines in emotion understanding, response quality, and audio-visual consistency. We open-source EmpaAva with an online live demo.

Figures

Figures reproduced from arXiv: 2608.04709 by the authors.

**Figure 2.** Workflow of the EmpaAva system. through empathetic planning, to embodied 3D expression. EmpaAva occupies exactly this space, uniting speech recognition, speech emotion recognition, LLM-based empathetic planning, structured reply plans, emotional speech synthesis, audiodriven facial motion, FLAME control, and 3DGS rendering in a single agentic system ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The Tri-Agent architecture of EmpaAva. PERCEPTIONAGENT understands the user, RESPONSEAGENT plans an empathetic reply, and RENDERAGENT turns the plan into an embodied 3D-avatar video. play. A lightweight registration is requested only when the user exports the full history. 4 Implementation Specification EmpaAva turns a stream of user audio-visual input into an embodied avatar reply that can be seen, heard, and answe… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Automatic evaluation on the AvaMERG test [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Qualitative multi-turn case studies of EmpaAva. rate 5 turns over four anonymized responses (50 preference votes); MECS and AV-Agree serve as cross-modal diagnostics (Appendix E). Results. In [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Qualitative avatar renderings across different [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 17 canonical work pages

[1]

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , year =

Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset , author =. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , year =

work page
[2]

2019 , pages =

Lin, Zhaojiang and Madotto, Andrea and Shin, Jamin and Xu, Peng and Fung, Pascale , booktitle =. 2019 , pages =

work page 2019
[3]

Majumder, Navonil and Hong, Pengfei and Peng, Shanshan and Lu, Jiankun and Ghosal, Deepanway and Gelbukh, Alexander and Mihalcea, Rada and Poria, Soujanya , journal =

work page
[4]

2020 , pages =

Li, Qintong and Chen, Hongshen and Ren, Zhaochun and Ren, Pengjie and Tu, Zhaopeng and Chen, Zhumin , booktitle =. 2020 , pages =

work page 2020
[5]

2019 , pages =

Poria, Soujanya and Hazarika, Devamanyu and Majumder, Navonil and Naik, Gautam and Cambria, Erik and Mihalcea, Rada , booktitle =. 2019 , pages =

work page 2019
[6]

2024 , pages =

Fei, Hao and Zhang, Han and Wang, Bin and Liao, Lizi and Liu, Qian and Cambria, Erik , booktitle =. 2024 , pages =

work page 2024
[7]

arXiv preprint arXiv:2502.04976 , year =

Towards Multimodal Empathetic Response Generation: A Rich Text-Speech-Vision Avatar-based Benchmark , author =. arXiv preprint arXiv:2502.04976 , year =

arXiv
[8]

Proceedings of the 28th ACM International Conference on Multimedia , year =

A Lip Sync Expert Is All You Need for Speech to Lip Generation in the Wild , author =. Proceedings of the 28th ACM International Conference on Multimedia , year =

work page

Show all 26 references

[9]

Zhang, Wenxuan and Cun, Xiaodong and Wang, Xuan and Zhang, Yong and Shen, Xi and Guo, Yu and Shan, Ying and Wang, Fei , journal =
[10]

Kim, Jisoo and Cho, Jungbin and Park, Joonho and Hwang, Soonmin and Kim, Da Eun and Kim, Geon and Yu, Youngjae , journal =
[11]

ACM Transactions on Graphics , volume =

Learning a Model of Facial Shape and Expression from 4D Scans , author =. ACM Transactions on Graphics , volume =. 2017 , pages =

2017
[12]

ACM Transactions on Graphics , volume =

3D Gaussian Splatting for Real-Time Radiance Field Rendering , author =. ACM Transactions on Graphics , volume =. 2023 , pages =

2023
[13]

arXiv preprint arXiv:2312.02069 , year =

Qian, Shenhan and Kirschstein, Tobias and Schoneveld, Liam and Davoli, Davide and Giebenhain, Simon and Nie. arXiv preprint arXiv:2312.02069 , year =

arXiv
[14]

arXiv preprint arXiv:2412.02508 , year =

When Words Smile: Generating Diverse Emotional Facial Expressions from Text , author =. arXiv preprint arXiv:2412.02508 , year =
[15]

Hu, Xiaolin and Yuan, Hang and Sang, Xinzhu and Yan, Binbin and Yu, Zhou and Huang, Cong and Chen, Kai , journal =
[16]

and Huang, Yanping and Dai, Andrew M

Chung, Hyung Won and Hou, Le and Longpre, Shayne and Zoph, Barret and Tay, Yi and Fedus, William and Li, Eric and Wang, Xuezhi and Dehghani, Mostafa and Brahma, Siddhartha and Webson, Albert and Gu, Shixiang Shane and Dai, Zhuyun and Suzgun, Mirac and Chen, Xinyun and Chowdher...

arXiv
[17]

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , year =

Towards Emotional Support Dialog Systems , author =. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , year =
[18]

Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =
[19]

arXiv preprint arXiv:2309.07864 , year =

The Rise and Potential of Large Language Model Based Agents: A Survey , author =. arXiv preprint arXiv:2309.07864 , year =

arXiv
[20]

Shen, Yongliang and Song, Kaitao and Tan, Xu and Li, Dongsheng and Lu, Weiming and Zhuang, Yueting , booktitle =
[21]

Tian, Linrui and Wang, Qi and Zhang, Bang and Bo, Liefeng , booktitle =
[22]

Xu, Sicheng and Chen, Guojun and Guo, Yu-Xiao and Yang, Jiaolong and Li, Chong and Zang, Zhenyu and Zhang, Yizhong and Tong, Xin and Guo, Baining , booktitle =
[23]

2024 , pages =

Cho, Kyusun and Lee, Joungbin and Yoon, Heeji and Hong, Yeobin and Ko, Jaehoon and Ahn, Sangjun and Kim, Seungryong , booktitle =. 2024 , pages =

2024
[24]

arXiv preprint arXiv:2406.08801 , year =

Hallo: Hierarchical Audio-Driven Visual Synthesis for Portrait Image Animation , author =. arXiv preprint arXiv:2406.08801 , year =

arXiv
[25]

2023 , howpublished =

2023
[26]

2020 , eprint=

Recipes for building an open-domain chatbot , author=. 2020 , eprint=

2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , year =

Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset , author =. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , year =

work page

[2] [2]

2019 , pages =

Lin, Zhaojiang and Madotto, Andrea and Shin, Jamin and Xu, Peng and Fung, Pascale , booktitle =. 2019 , pages =

work page 2019

[3] [3]

Majumder, Navonil and Hong, Pengfei and Peng, Shanshan and Lu, Jiankun and Ghosal, Deepanway and Gelbukh, Alexander and Mihalcea, Rada and Poria, Soujanya , journal =

work page

[4] [4]

2020 , pages =

Li, Qintong and Chen, Hongshen and Ren, Zhaochun and Ren, Pengjie and Tu, Zhaopeng and Chen, Zhumin , booktitle =. 2020 , pages =

work page 2020

[5] [5]

2019 , pages =

Poria, Soujanya and Hazarika, Devamanyu and Majumder, Navonil and Naik, Gautam and Cambria, Erik and Mihalcea, Rada , booktitle =. 2019 , pages =

work page 2019

[6] [6]

2024 , pages =

Fei, Hao and Zhang, Han and Wang, Bin and Liao, Lizi and Liu, Qian and Cambria, Erik , booktitle =. 2024 , pages =

work page 2024

[7] [7]

arXiv preprint arXiv:2502.04976 , year =

Towards Multimodal Empathetic Response Generation: A Rich Text-Speech-Vision Avatar-based Benchmark , author =. arXiv preprint arXiv:2502.04976 , year =

arXiv

[8] [8]

Proceedings of the 28th ACM International Conference on Multimedia , year =

A Lip Sync Expert Is All You Need for Speech to Lip Generation in the Wild , author =. Proceedings of the 28th ACM International Conference on Multimedia , year =

work page