JoyAI-Talker: Full-Duplex Speech Interactive Large Model Built for Empathetic Voice Agents

Boya Dong; Chao Xue; Fan Yu; Hao Li; Hongfei Xue; Jiaqi Wang; Ji Miao; Jingdong Li; Jinming Chen; Lin Zhu

REVIEW 4 major objections 6 minor 19 references

JoyAI-Talker: Full-Duplex Speech Interactive Large Model Built for Empathetic Voice Agents

T0 review · 4 major / 6 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 JoyAI-Talker claims a full-duplex voice agent can keep its text-model reasoning—94.62% on MATH and 88.57% on MMLU—by training speech and text jointly from the mid-training stage onward, while a separate gating module handles turn-taking.

desk verdict Solid industrial report, but the key claim—joint training preserves text cognition—lacks the backbone baseline needed to support it. read the letter →

arxiv 2608.01119 v1 pith:FRAJGL7C submitted 2026-08-02 cs.SD

Yinhao Bai , Jinming Chen , Yafeng Chen , Wei Deng , Boya Dong , Nan Duan , Yu Gu , Weisheng Han

show 21 more authors

Yankun Huang Ming Ke Hao Li Jingdong Li Xiangyu Liang Ning Liu Yuan Liu Ji Miao Jiaqi Wang Qi Wang Wenchao Wang Yuxuan Wang Zhenfang Wang Zhangyu Xiao Chao Xue Hongfei Xue Fan Yu Tianyi Zhang Yuan Zhang Yuqi Zhang Lin Zhu

This is my paper

classification cs.SD

keywords full-duplexspeechdialoguecognitivedegradationspeech-textjointtrainingempathychain-of-thoughtparalinguisticcontrolMixture-of-Expertsturn-takinggating

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

JoyAI-Talker is a full-duplex speech dialogue system built to answer one question: can a voice agent keep the reasoning ability of a large text model once it learns to listen and speak? The paper's answer is yes, provided speech and text are trained together from the mid-training phase onward rather than added at the end. The authors report that the final model scores 94.62% on MATH, 88.57% on MMLU, and 64.51% on GPQA in text, while also doing competitive speech recognition, translation, question answering, and emotion perception from audio. Turn-taking is handled by a separate state-driven gating module that responds to 0.88 of user interruptions with low false triggers, so the language core is not burdened with real-time streaming decisions. A persona-adaptive response pipeline extracts age, gender, and emotion from the user's voice and feeds those cues into chain-of-thought reasoning and expressive speech control.

What carries the argument

The load-bearing machinery is the decoupled Duplex-Thinker-Talker architecture combined with a unified speech-text joint training pipeline. The pipeline interleaves text and speech data at every post-pretraining stage with a strict loss mask that computes the autoregressive loss only on assistant responses; the decoupling routes real-time turn-taking decisions—encoded as state tokens such as <|partial|>, <|complete|>, <|accept|>, <|reject|>, and <|backchannel|>—through a small external gating decoder, so high-frequency streaming signals do not fragment the attention or parameter capacity of the language core.

What would settle it

Run the paper's exact text-to-text evaluation harness on the leading open-weight speech-language baseline it compares against; if the baseline's MATH, MMLU, or GPQA scores rise to or above JoyAI-Talker's reported numbers, the claimed preservation is an artifact of protocol. Alternatively, train the same backbone with speech data only during supervised fine-tuning (no early joint mid-training) and compare MATH/MMLU; equal scores would show the timing of joint training is not what preserves cognition.

Watch

Extended reading notes

Core claim

The central claim is that cognitive degradation—the usual drop in text reasoning when a model learns speech—can be alleviated by a unified speech-text joint training pipeline applied across mid-training, context extension, supervised fine-tuning, and preference optimization, and by a decoupled Duplex-Thinker-Talker architecture. The Thinker (a 48.9B-parameter Mixture-of-Experts language model) handles understanding and response planning; the Talker turns text, natural-language instructions, and paralinguistic tokens into expressive speech; a separate 1.7B-parameter Duplex decoder handles turn-taking via state tokens. Evidence: the final checkpoint scores 94.62% on MATH, 88.57% on MMLU, and 6

Load-bearing premise

The central claim rests on the assumption that the model's benchmark scores, measured under its own chat-template greedy-decoding protocol, are directly comparable to the published numbers of the leading baseline it beats, and that the preserved cognitive ability is caused by the joint-training design rather than by evaluation differences or other training choices.

Editorial extensions

If this is right

Voice agents built on this design can answer math, logic, and knowledge questions from audio without falling back to a text-only model.
Full-duplex conversation can be added to an existing half-duplex speech model as a plug-in gate, without retraining the main backbone or degrading its reasoning.
Explicitly inserting perceived speaker attributes (age, gender, emotion) into chain-of-thought reasoning improves both the content and the acoustic expression of empathetic responses.
Speech-only preference optimization, applied within the speech-conditioned pool, improves instruction following and reduces hallucinations without cross-modal distribution mismatch.
Raw audio can drive agentic tool calling, with the model transferring text-domain function-calling competence to speech intents.

Reading between the lines

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

Editorial inference: the causal claim that early joint training prevents cognitive degradation is not directly demonstrated, because no ablation freezes the backbone or moves speech training to supervised fine-tuning only; the reported tables compare final checkpoints, so other training choices could carry the preservation.
Editorial inference: if the preservation result is real, a natural extension is to push speech-text joint training into earlier pretraining stages or vary the speech-to-text data ratio to map the trade-off curve.
Editorial inference: the Joy-Duplex results suggest turn-taking competence is largely separable from language intelligence; a direct test would swap the 1.7B gating decoder for larger or smaller decoders and see whether full-duplex behavior changes independently of text benchmarks.
Editorial inference: PAER's chain-of-thought format is testable against a no-reasoning variant; if removing the explicit attribute-generation step drops empathy scores more than removing paralinguistic tokens, that would isolate where the empathetic gain comes from.

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

You should read this if you care about speech LLMs: it's a credible industrial report from JD.com with a clean modular design and a lot of benchmark tables. The headline claim is that their speech-text joint training from mid-training onward avoids the usual cognitive degradation, and they show high MATH/MMLU/GPQA scores. But the evidence for that claim isn't in the paper: Table 1 compares JoyAI-Talker SFT/DPO against Qwen3-Omni, with no JoyAI-LLM Flash column and no frozen-backbone or text-only control. So we can't tell whether the joint training preserved anything, or whether the backbone was simply strong to begin with. The stress-test note is right—the missing baseline is load-bearing, and the protocol mismatch with Qwen3-Omni's published numbers is a secondary issue by comparison.

What's genuinely good: the architecture description is detailed and honest. The decoder-only duplex gating with state tokens (partial/complete/backchannel/accept/reject) is a sensible plug-and-play design, and the duplex results on Full-Duplex-Bench v1.5 are strong, though the Gemini 3.1 Live API comparison is under-specified. The Talker's voice-conversion-based SFT data construction is a practical trick. The PAER pipeline is well motivated, and the paper openly admits limitations and cites prior work without exaggerating novelty.

Soft spots, in proportion: (1) The missing backbone baseline—without it, the causal claim about joint training is asserted, not shown. (2) Evaluation comparability: they use chat-format OpenCompass with greedy decoding, while Qwen3-Omni's published numbers may come from a different protocol; the paper doesn't pin that down. (3) The MER2025 label refinement in §3.3.1 changes the test set, so the emotion numbers aren't directly comparable to published ones; the paper discloses this but doesn't quantify the effect. (4) The LLM-as-a-judge empathy scores have no judge details or human agreement—suggestive at best. (5) No code or weights, so no independent reproduction.

Who this is for: engineers building speech agents will get the most value; researchers studying cognitive degradation in speech-text training should treat this as a strong prompt to run the missing ablation. It deserves a serious referee, but with a required revision: add the backbone baseline and a text-only control, pin down the Gemini and MER2025 protocols, and release at least some artifacts. I'd send it to peer review, not accept it as is.

Referee Report

4 major / 6 minor

Summary. The paper introduces JoyAI-Talker, a full-duplex speech dialogue system built on a decoupled Duplex-Thinker-Talker architecture. The Thinker is a 48.9B sparse-MoE LLM (JoyAI-LLM Flash) trained with a unified speech-text joint pipeline spanning mid-training, context extension, SFT, and DPO. The Talker is an instruction-controllable expressive TTS module derived from JoyVoice, and Joy-Duplex is a state-driven, plug-and-play full-duplex gating engine. The paper reports competitive T2T and S2T benchmark numbers, empathetic response improvements on EchoMind, and strong Full-Duplex-Bench v1.5 results (interruption response 0.88, low false triggers). The central claim is that the joint training pipeline 'largely preserves' the backbone's text reasoning and STEM capabilities while adding speech abilities.

Significance. If the reported results hold, the paper makes a useful engineering contribution: it demonstrates that a modular Thinker-Talker design with early, phased speech-text joint training can yield a speech dialogue model with strong text-domain reasoning (e.g., 94.62 on MATH, 88.57 on MMLU, 64.51 on GPQA) and competitive full-duplex behavior. The use of external T2T benchmarks as a negative control is a good methodological instinct, and the Full-Duplex-Bench evaluation covers a realistic overlap taxonomy. The PAER framework and instruction-controllable Talker address a genuinely important problem: making empathetic voice agents that preserve speaker identity while producing paralinguistic events. However, the paper's headline causal claim—that the joint-training design specifically prevents cognitive degradation—is not yet supported by the evidence as presented.

major comments (4)

[§3.1, Table 1]
[§3.1, evaluation protocol]
[§3.4.2, Table 5]
[§3.3.1, Table 3]

minor comments (6)

[Abstract]
[Table 1]
[§2.1.1, DPO]
[§3.5]
[§2.3]
[§3.4.2]

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; key benchmark claims rest on external, non-fitted evaluations, while the main evidential gap (missing backbone control) is an omitted-control issue, not circularity.

full rationale

I walked the claimed derivation chain. The central T2T 'cognitive preservation' result (MATH 94.62%, MMLU 88.57%, GPQA 64.51%) is measured on external benchmarks under OpenCompass; no parameter of the model is fitted to these test sets, so the scores are not equivalent to an input by construction. The S2T and EQ results are likewise external benchmarks (Aishell, CoVoST, OpenBookQA, AIR-Bench, MER2025), and the full-duplex results use Full-Duplex-Bench v1.5, an external protocol. The self-citations to JoyAI-LLM Flash [5] and JoyVoice [8] provide architectural provenance and backbone initialization; they do not function as a uniqueness theorem and are not the evidence for the benchmark numbers. Two concerns were flagged but do not constitute circularity. First, §3.1/Table 1 says it will 'examine whether JoyAI-Talker preserves the text-domain capabilities of its backbone' but reports no JoyAI-LLM Flash column or frozen-backbone control; this is an omitted control that weakens the causal 'preservation' attribution, not a reduction of the result to its inputs. Second, §3.3.1 states the authors 'audited the test set labels' of MER2025 and evaluated emotion on 'our refined MER2025 annotations'; this changes protocol comparability with Qwen3-Omni's published number, but there is no statement that the refined labels were used in training, so it is a benchmark-integrity issue, not a self-consistency loop. No fitted input is renamed as a prediction, and no equation reduces to a definition. Score 0.

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

The paper introduces no new physical or mathematical entity. Its free parameters are effectively the data mixtures, training-stage proportions, judge rubrics, and duplex state thresholds, none of which are quantified in the text. The assumptions above are the load-bearing premises: the fairness of the benchmark protocol, the causal power of the joint-training pipeline without ablations, the reliability of LLM-judge empathy scores, and the fidelity of the Gemini API evaluation.

assumptions (4)

domain assumption The benchmark protocol used to produce the author-reported T2T and S2T scores, especially for Qwen3-Omni comparisons, is comparable without undisclosed tuning.
'All models evaluated in this section are tested with chat-format inference templates, including conversational special tokens' (§3.1). The paper does not describe the prompt templates, few-shot settings, or post-processing for each benchmark or baseline, yet the comparison rests on these being fair.
domain assumption The early joint training pipeline, described in §2.1.1, is the cause of the preserved text scores, and a frozen-backbone baseline or a speech-only-SFT baseline would show worse performance.
No ablation is reported. The claim in §2.1.1 that integrating the speech modality at an earlier stage prevents catastrophic forgetting is asserted, not demonstrated. Without an ablation, the text scores could be explained by the backbone's strength alone.
domain assumption LLM-as-a-judge ratings in §3.3.2 reflect genuine empathy and response quality.
The six weighted dimensions are defined, but no judge model, prompt, temperature, agreement with human raters, or bias analysis is reported. The high scores (8-9 out of 10) therefore rest on unverified judge reliability.
domain assumption The reported Full-Duplex-Bench v1.5 protocol and the Gemini 3.1 Live API evaluation are equivalent to the original benchmark's conditions.
In §3.4.2 the authors state they evaluated Gemini 3.1 Live via a commercial API, but give no API version, date, latency, or classification details. The comparison assumes the conditions match the benchmark paper's protocol.

how reviews work

0 comments

Cite this review

Pith. "Pith review of JoyAI-Talker: Full-Duplex Speech Interactive Large Model Built for Empathetic Voice Agents." pith.science (2026). https://pith.science/paper/FRAJGL7C

@misc{pith2026260801119,
  author       = {Pith},
  title        = {Pith review of: JoyAI-Talker: Full-Duplex Speech Interactive Large Model Built for Empathetic Voice Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FRAJGL7C}},
  note         = {Machine review of arXiv:2608.01119}
}

read the original abstract

We present JoyAI-Talker, a full-duplex speech dialogue system that delivers robust foundation model capabilities while empowering empathetic interaction and voice agent intelligence. JoyAI-Talker adopts a modular Thinker-Talker architecture and further implements a unified speech-text joint training pipeline to mitigate the common "cognitive degradation" bottleneck, thereby largely preserving the model's core textual reasoning, STEM, and logical capabilities while extending them to speech-based interaction. For expressive speech synthesis, the Talker module employs a text-controllable generation paradigm that enables natural-language instructions to flexibly control vocal attributes and localized paralinguistic events, such as laughter and sighs, supporting more expressive and fine-grained speech responses. To enhance conversational empathy, we introduce the Persona-Adaptive Empathetic Response (PAER) framework. PAER employs a hierarchical cognitive pipeline to extract non-verbal speaker cues, such as gender, age, and emotional state, from raw input audio, incorporate them into the Thinker's CoT reasoning, and generate context-adaptive responses that align semantically appropriate text with fine-grained control over utterance-level expressiveness and localized paralinguistic events, including sighs, speaking rate, and volume. We further integrate Joy-Duplex, a state-driven, plug-and-play full-duplex framework that functions as an efficient gating engine for real-time turn control. Extensive evaluations show that JoyAI-Talker achieves highly competitive performance on foundational T2T and S2T benchmarks. In full-duplex evaluation, the system reaches a high response rate of 0.88 under user interruptions while maintaining an extremely low false-trigger rate under background speech, demonstrating its readiness for fluid and natural speech dialogue.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 10 canonical work pages

[1]

Moshi: a speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037, 2024

Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, and Neil Zeghidour. Moshi: a speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037, 2024

arXiv 2024
[2]

Freeze-omni: A smart and low latency speech-to-speech dialogue model with frozen llm.arXiv preprint arXiv:2411.00774, 2024

VITA Team. Freeze-omni: A smart and low latency speech-to-speech dialogue model with frozen llm.arXiv preprint arXiv:2411.00774, 2024

arXiv 2024
[3]

Qwen3-omni technical report.arXiv preprint arXiv:2509.17765, 2025

Qwen Team. Qwen3-omni technical report.arXiv preprint arXiv:2509.17765, 2025

arXiv 2025
[4]

Full-duplex-bench v1

Guan-Ting Lin, Shih-Yun Shan Kuan, Qirui Wang, Jiachen Lian, Tingle Li, Shinji Watanabe, and Hung-yi Lee. Full-duplex-bench v1. 5: Evaluating overlap handling for full-duplex speech models. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 19447–19451. IEEE, 2026

work page 2026
[5]

Joyai-llm flash: Advancing mid-scale llms with token efficiency, 2026

Aichen Cai, Anmeng Zhang, Anyu Li, Bo Zhang, Bohua Cai, Chang Li, Changjian Jiang, Changkai Lu, Chao Xue, et al. Joyai-llm flash: Advancing mid-scale llms with token efficiency, 2026

work page 2026
[6]

Deepseek-v3 technical report, 2025

Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report, 2025

work page 2025
[7]

Charles, Cheng Chen, Guanduo Chen, Haiting Chen, Huarong Chen, Jiahao Chen, et al

Kimi Team, Yifan Bai, Yiping Bao, Y. Charles, Cheng Chen, Guanduo Chen, Haiting Chen, Huarong Chen, Jiahao Chen, et al. Kimi k2: Open agentic intelligence, 2026

work page 2026
[8]

Joyvoice: Long-context conditioning for anthropomorphic multi-speaker conversational synthesis

Speech Team, JD. Joyvoice: Long-context conditioning for anthropomorphic multi-speaker conversational synthesis. arXiv preprint arXiv:2512.19090, 2025

arXiv 2025

Show all 19 references

[9]

Interaction models: A scalable approach to human-ai collaboration.Thinking Machines Lab: Connectionism, May 2026

Thinking Machines Lab. Interaction models: A scalable approach to human-ai collaboration.Thinking Machines Lab: Connectionism, May 2026. https://thinkingmachines.ai/blog/interaction-models/

2026
[10]

Bin Wang, Xunlong Zou, Geyu Lin, Shuo Sun, Zhuohan Liu, Wenyu Zhang, Zhengyuan Liu, AiTi Aw, and Nancy F. Chen. Audiobench: A universal benchmark for audio large language models. InProceedings of the 2025 19 Conference of the Nations of the Americas Chapter of the Association ...

2025
[11]

Xize Cheng, Ruofan Hu, Xiaoda Yang, Jingyu Lu, Dongjie Fu, Zehan Wang, Shengpeng Ji, Rongjie Huang, Boyang Zhang, Tao Jin, and Zhou Zhao. Voxdialogue: Can spoken dialogue systems understand information beyond words? InThe Thirteenth International Conference on Learning Represe...

2025
[12]

Uro-bench: Towards comprehensive evaluation for end-to-end spoken dialogue models

Ruiqi Yan, Xiquan Li, Wenxi Chen, Zhikang Niu, Chen Yang, Ziyang Ma, Kai Yu, and Xie Chen. Uro-bench: Towards comprehensive evaluation for end-to-end spoken dialogue models. InFindings of the Association for Computational Linguistics: EMNLP 2025, Suzhou, China, November 4-9, 2...

2025
[13]

Air-bench: Benchmarking large audio-language models via generative comprehension

Qian Yang, Jin Xu, Wenrui Liu, Yunfei Chu, Ziyue Jiang, Xiaohuan Zhou, Yichong Leng, Yuanjun Lv, Zhou Zhao, Chang Zhou, and Jingren Zhou. Air-bench: Benchmarking large audio-language models via generative comprehension. In Proceedings of the 62nd Annual Meeting of the Associat...

2024
[14]

Schuller, and Jianhua Tao

Zheng Lian, Rui Liu, Kele Xu, Bin Liu, Xuefei Liu, Yazhou Zhang, Xin Liu, Yong Li, Zebang Cheng, Haolin Zuo, Ziyang Ma, Xiaojiang Peng, Xie Chen, Ya Li, Erik Cambria, Guoying Zhao, Björn W. Schuller, and Jianhua Tao. MER 2025: When affective computing meets large language mode...

2025
[15]

Echomind: An interrelated multi-level benchmark for evaluating empathetic speech language models.CoRR, abs/2510.22758, 2025

Li Zhou, Lutong Yu, You Lyu, Yihang Lin, Zefeng Zhao, Junyi Ao, Yuhao Zhang, Benyou Wang, and Haizhou Li. Echomind: An interrelated multi-level benchmark for evaluating empathetic speech language models.CoRR, abs/2510.22758, 2025

2025
[16]

Easy turn: Integrating acoustic and linguistic modalities for robust turn-taking in full-duplex spoken dialogue systems.CoRR, abs/2509.23938, 2025

Guojian Li, Chengyou Wang, Hongfei Xue, Shuiyuan Wang, Dehui Gao, Zihan Zhang, Yuke Lin, Wenjie Li, Longshuai Xiao, Zhonghua Fu, and Lei Xie. Easy turn: Integrating acoustic and linguistic modalities for robust turn-taking in full-duplex spoken dialogue systems.CoRR, abs/2509....

2025
[17]

Soulx-duplug: Plug-and-play streaming state prediction module for realtime full-duplex speech conversation.arXiv preprint arXiv:2603.14877, 2026

Ruiqi Yan, Wenxi Chen, Zhanxun Liu, Ziyang Ma, Haopeng Lin, Hanlin Wen, Hanke Xie, Jun Wu, Yuzhe Liang, Yuxiang Zhao, et al. Soulx-duplug: Plug-and-play streaming state prediction module for realtime full-duplex speech conversation.arXiv preprint arXiv:2603.14877, 2026

2026
[18]

Fastturn: Unifying acoustic and streaming semantic cues for low-latency and robust turn detection.CoRR, abs/2604.01897, 2026

Chengyou Wang, Hongfei Xue, Chunjiang He, Jingbin Hu, Shuiyuan Wang, Bo Wu, Yuyu Ji, Jimeng Zheng, Ruofei Chen, Zhou Zhu, and Lei Xie. Fastturn: Unifying acoustic and streaming semantic cues for low-latency and robust turn detection.CoRR, abs/2604.01897, 2026

2026 arXiv
[19]

Unified streaming and non-streaming two-pass end-to-end model for speech recognition.arXiv preprint arXiv:2012.05481, 2020

Binbin Zhang, Di Wu, Zhuoyuan Yao, Xiong Wang, Fan Yu, Chao Yang, Liyong Guo, Yaguang Hu, Lei Xie, and Xin Lei. Unified streaming and non-streaming two-pass end-to-end model for speech recognition.arXiv preprint arXiv:2012.05481, 2020. 20

2012 arXiv

Pith tools

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

[1] [1]

Moshi: a speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037, 2024

Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, and Neil Zeghidour. Moshi: a speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037, 2024

arXiv 2024

[2] [2]

Freeze-omni: A smart and low latency speech-to-speech dialogue model with frozen llm.arXiv preprint arXiv:2411.00774, 2024

VITA Team. Freeze-omni: A smart and low latency speech-to-speech dialogue model with frozen llm.arXiv preprint arXiv:2411.00774, 2024

arXiv 2024

[3] [3]

Qwen3-omni technical report.arXiv preprint arXiv:2509.17765, 2025

Qwen Team. Qwen3-omni technical report.arXiv preprint arXiv:2509.17765, 2025

arXiv 2025

[4] [4]

Full-duplex-bench v1

Guan-Ting Lin, Shih-Yun Shan Kuan, Qirui Wang, Jiachen Lian, Tingle Li, Shinji Watanabe, and Hung-yi Lee. Full-duplex-bench v1. 5: Evaluating overlap handling for full-duplex speech models. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 19447–19451. IEEE, 2026

work page 2026

[5] [5]

Joyai-llm flash: Advancing mid-scale llms with token efficiency, 2026

Aichen Cai, Anmeng Zhang, Anyu Li, Bo Zhang, Bohua Cai, Chang Li, Changjian Jiang, Changkai Lu, Chao Xue, et al. Joyai-llm flash: Advancing mid-scale llms with token efficiency, 2026

work page 2026

[6] [6]

Deepseek-v3 technical report, 2025

Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report, 2025

work page 2025

[7] [7]

Charles, Cheng Chen, Guanduo Chen, Haiting Chen, Huarong Chen, Jiahao Chen, et al

Kimi Team, Yifan Bai, Yiping Bao, Y. Charles, Cheng Chen, Guanduo Chen, Haiting Chen, Huarong Chen, Jiahao Chen, et al. Kimi k2: Open agentic intelligence, 2026

work page 2026

[8] [8]

Joyvoice: Long-context conditioning for anthropomorphic multi-speaker conversational synthesis

Speech Team, JD. Joyvoice: Long-context conditioning for anthropomorphic multi-speaker conversational synthesis. arXiv preprint arXiv:2512.19090, 2025

arXiv 2025