Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Towards a Japanese Full-duplex Spoken Dialogue System

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read J-Moshi, the first publicly available Japanese full-duplex spoken dialogue model, outperforms the dGSLM baseline in both naturalness and meaningfulness.

desk verdict First public Japanese full-duplex dialogue model with released artifacts; evaluation has a matched-prompt question, but the contribution is solid and worth refereeing. read the letter →

arxiv 2506.02979 v1 pith:A25F2PUD submitted 2025-06-03 cs.CL eess.AS

Atsumoto Ohashi , Shinya Iizuka , Jingjing Jiang , Ryuichiro Higashinaka This is my paper
classification cs.CLeess.AS
keywords full-duplexspokendialogueJapaneseMoshiRQ-TransformerJ-CHATmulti-streamTTSspeechoverlapbackchannel
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 reports the first publicly available full-duplex spoken dialogue model for Japanese, built by adapting the English model Moshi. It claims that the adapted model, J-Moshi, and its TTS-augmented extension J-Moshi-ext, generate dialogue continuations that human raters find more natural and more meaningful than those of dGSLM, a Japanese-trained baseline. The authors also report that J-Moshi reproduces Japanese-specific turn-taking patterns, notably more speech overlaps than English Moshi. If the claims hold, this gives the Japanese speech community a working baseline and a path for porting full-duplex dialogue capabilities to other languages.

What carries the argument

The central mechanism is the RQ-Transformer from Moshi, a 7B-parameter Temporal Transformer plus a smaller Depth Transformer that autoregressively models 17 parallel token streams: one text stream for Moshi's inner monologue, eight audio streams for Moshi's speech, and eight for the user's speech, at 12.5 Hz. The Mimi neural codec encodes speech into semantic and acoustic tokens. Training proceeds in two stages: pre-training on J-CHAT to learn Japanese spoken dialogue, then fine-tuning on stereo data to learn overlap and backchannel timing; a multi-stream TTS stage synthesizes additional stereo dialogues and improves language quality. A temperature of τ=0.8 is used for generation.

What would settle it

Run a matched human evaluation in which all models continue the same 50 ten-second prompts, keeping temperature and evaluator instructions identical. If J-Moshi and J-Moshi-ext no longer beat dGSLM in naturalness and meaningfulness, the paper's central comparison fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that Moshi's architecture can be transferred to Japanese by swapping the text tokenizer to a Japanese SentencePiece model, freezing the Mimi audio codec, pre-training the RQ-Transformer on 69,000 hours of J-CHAT, and then fine-tuning on 344 hours of stereo spoken dialogue. The fine-tuned J-Moshi scores 2.67 for naturalness and 2.19 for meaningfulness on a 5-point scale, versus 2.44 and 1.76 for dGSLM. Adding 602 hours of multi-stream TTS synthetic dialogue, J-Moshi-ext, raises meaningfulness to 2.30. The paper also shows higher overlap and IPU counts than English Moshi, consistent with Japanese conversational norms.

Load-bearing premise

The human evaluation compares models on 50 samples randomly selected from each model's 709 outputs, without stating that the same 50 prompts were used for every model; if prompts differ, score gaps could reflect prompt difficulty rather than model quality.

Editorial extensions

If this is right

  • J-Moshi and J-Moshi-ext are released publicly, giving researchers a Japanese full-duplex baseline to compare against.
  • Multi-stream TTS augmentation raises meaningfulness from 2.19 to 2.30, so synthetic stereo data is a viable route to improve language capability when real stereo corpora are scarce.
  • The model reproduces Japanese-specific turn-taking, with more overlaps than English Moshi, suggesting full-duplex behavior is learnable from the target language's data.
  • The 88% PAD token ratio in J-CHAT indicates Japanese text streams are sparser relative to audio, pointing to a need for language-specific loss weighting in future adaptations.

Reading between the lines

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

  • A natural next test is whether the same two-stage recipe transfers to other languages with large monophonic corpora but little stereo dialogue, such as Mandarin or Korean.
  • Because Mimi was frozen and resynthesis scores remain about 0.5 points below ground truth, adapting the codec to Japanese acoustic patterns may be a higher-yield improvement than scaling the language model alone.
  • The human evaluation's random selection of 50 samples per model, without a stated shared prompt set, leaves open the possibility that part of the score gap reflects prompt difficulty; a matched-prompt evaluation would settle that.
  • The high PAD ratio suggests that a tokenizer tuned for Japanese could reduce the 88% PAD rate, potentially improving both training efficiency and fluency.
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 / 5 minor

Summary. The paper adapts Moshi, an English full-duplex spoken dialogue model, to Japanese by replacing the text tokenizer with a Japanese SentencePiece model, randomly initializing the vocabulary-related weights, pre-training the RQ-Transformer on about 60,000 hours of J-CHAT data, and fine-tuning on 344 hours of stereo spoken dialogue plus 602 hours of multi-stream TTS synthetic dialogue. The resulting models, J-Moshi and J-Moshi-ext, are evaluated on a prompted dialogue continuation task against a Japanese dGSLM baseline, with automatic perplexity and human naturalness/meaningfulness ratings. The paper claims that the proposed models outperform the Japanese baseline in both metrics and that they acquire Japanese-specific turn-taking behavior, evidenced by more overlaps relative to English Moshi.

Significance. If the central claims hold, this is a useful and timely contribution: it reports the first publicly available Japanese full-duplex spoken dialogue model, releases code, models, and speech samples, and provides a detailed recipe for adapting a large English full-duplex model to another language. The paper also gives concrete evidence that Mimi can encode and resynthesize Japanese without retraining, and it offers preliminary cross-linguistic turn-taking observations that could inform future multilingual full-duplex research. The strengths include reproducible training configurations (loss weights, learning rates, batch sizes, hardware), transparent data handling, and comparison against a trained-from-scratch dGSLM baseline. The main weakness is that the human evaluation protocol leaves open a prompt-difficulty confound, and the reported statistical support is incomplete; these issues directly affect the headline claim.

major comments (3)
  1. [4.3, Table 2] The human evaluation sampling procedure is not sufficiently specified. The text says 50 audio samples were 'randomly selected from the 709 dialogue continuations of each model,' which can be read as independent selection per model. If the 50 prompts differ across dGSLM, J-Moshi, J-Moshi-ext, Re-synthesis, and Ground-truth, then the naturalness and meaningfulness differences in Table 2 may partly reflect prompt difficulty rather than model quality. Since the naturalness advantage of J-Moshi over dGSLM is only 0.23 on a 5-point scale (2.67 vs. 2.44), this confound is load-bearing for the central claim. Please state explicitly whether the same 50 prompts were used for all conditions; if not, re-run the evaluation with matched prompts or otherwise control for prompt difficulty.
  2. [4.3, Table 2] No significance test is reported for the human evaluation comparisons. The text claims that J-Moshi and J-Moshi-ext 'significantly surpassed dGSLM in meaningfulness,' but Table 2 only lists per-condition 95% confidence intervals. With a naturalness difference of 0.23 and no paired test, the claim that the models outperform the baseline in both naturalness and meaningfulness is not statistically supported. Please report a paired significance test (e.g., Wilcoxon signed-rank or bootstrap over rated samples) and, ideally, the distribution of per-item score differences between conditions.
  3. [5, Table 3] The cross-linguistic turn-taking comparison uses values for English Moshi reported by Défossez et al. [6] on 1,000 different samples, whereas the Japanese values come from 709 prompts from a different corpus. The authors acknowledge that 'the experimental conditions are not strictly identical' but still conclude that J-Moshi has acquired Japanese-specific behavior such as more overlaps. This is a weaker form of evidence than a matched comparison would be; the observed differences could reflect prompt content, corpus genre, or annotation differences rather than language-specific behavior. Please either temper the conclusion or provide a matched analysis, such as running English Moshi on the same prompts or using a comparable diarization/annotation pipeline.
minor comments (5)
  1. [4.2, Figure 3] Figure 3 reports perplexity but no error bars or measures of variance, making it difficult to assess whether the approximately 100-point improvement over dGSLM at τ=0.8 is stable across random seeds or test chunks.
  2. [3.2] The heading 'J-CHA T Preprocessing' contains a typographical error; it should read 'J-CHAT Preprocessing.'
  3. [4.3] The paper does not state how the crowd-workers were assigned to conditions or whether each sample was rated by multiple workers; reporting per-sample mean scores and rater counts would improve transparency.
  4. [5] In Table 3, the four turn-taking metrics are presented as seconds per minute, but their relationship is not fully defined (e.g., whether Overlap is fully contained within IPU or counted separately); a brief definition or reference for the exact computation would help readers interpret the sums.
  5. [3.4] The selection of synthetic speech by lowest WER is reasonable, but the overall WER of 24.6% is relatively high; a sentence noting the implications for training data quality and possible filtering effects would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central evaluation compares independently trained models on held-out prompts with disclosed hyperparameters.

full rationale

The paper's derivation chain is empirical rather than definitional. J-Moshi and J-Moshi-ext are produced by adapting Moshi through pre-training on J-CHAT, fine-tuning on stereo Japanese dialogue corpora, and optional augmentation with multi-stream TTS synthetic data. The central claim of outperforming dGSLM rests on human evaluation of generated continuations plus automatic perplexity measures. No parameter is fitted to a target comparison: hyperparameters such as loss weights, temperature values, and learning rates are stated and were not chosen to force Table 2 outcomes. The baseline dGSLM is an independently established architecture trained by the authors on the same Japanese data, and Re-synthesis and Ground-truth provide external anchors. Self-citations in the reference list point to data sources (RealPersonaChat, Travel Agency Dialogue Corpus) and are not load-bearing theoretical justifications. No uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via citation; the Moshi architecture is adopted from an external technical report. The skeptic's concern that the 50 human-evaluated samples may have been selected independently per model is a methodological limitation about prompt matching, not a circularity: the evaluation is still an external comparison against held-out prompts and untrained baselines. Therefore the paper is self-contained against its benchmarks and receives a score of 0.

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

The central claim rests on transfer from Moshi's English pretrained backbone, on automatic speech processing quality for Japanese, and on the interpretation of turn-taking statistics. No new entities are postulated. The training hyperparameters (loss weights, temperatures, token delays) are standard and disclosed, not fitted constants; they do not enter a derivation that needs independent benchmarking, so the free-parameter list is empty.

assumptions (3)
  • domain assumption Moshi's pretrained 7B English LLM provides a transferable linguistic backbone, so only text token embeddings and Text Linear parameters need to be randomized.
    Section 3.1: The authors replace the tokenizer and initialize text-related weights randomly while keeping the rest of the Temporal Transformer, implicitly assuming English internal representations are useful for Japanese after further pre-training on 69,000 hours of J-CHAT.
  • domain assumption The diarization and ASR pipeline on J-CHAT yields channel separation and transcriptions accurate enough to pre-train a stereo dialogue model.
    Section 3.2: They process 69,000 hours with pyannote diarization and Whisper/WhisperX ASR without a stated quality filter, relying on these tools to assign speakers to channels and provide timestamped text tokens for pre-training.
  • domain assumption Japanese conversation exhibits more overlaps and backchannels than English, so the higher overlap rate in J-Moshi compared to Moshi is interpreted as having acquired Japanese-specific behavior.
    Section 5 and references [11,12]: The comparison in Table 3 assumes that the observed differences in IPU, Pause, Gap, and Overlap between Japanese generated dialogue and English generated dialogue reflect learned language-specific turn-taking, not artifacts of the different evaluation conditions (the Moshi values are taken from [6]).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards a Japanese Full-duplex Spoken Dialogue System." pith.science (2026). https://pith.science/paper/A25F2PUD

@misc{pith2026250602979,
  author       = {Pith},
  title        = {Pith review of: Towards a Japanese Full-duplex Spoken Dialogue System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A25F2PUD}},
  note         = {Machine review of arXiv:2506.02979}
}
read the original abstract

Full-duplex spoken dialogue systems, which can model simultaneous bidirectional features of human conversations such as speech overlaps and backchannels, have attracted significant attention recently. However, the study of full-duplex spoken dialogue systems for the Japanese language has been limited, and the research on their development in Japanese remains scarce. In this paper, we present the first publicly available full-duplex spoken dialogue model in Japanese, which is built upon Moshi, a full-duplex dialogue model in English. Our model is trained through a two-stage process: pre-training on a large-scale spoken dialogue data in Japanese, followed by fine-tuning on high-quality stereo spoken dialogue data. We further enhance the model's performance by incorporating synthetic dialogue data generated by a multi-stream text-to-speech system. Evaluation experiments demonstrate that the trained model outperforms Japanese baseline models in both naturalness and meaningfulness.

Figures

Figures reproduced from arXiv: 2506.02979 by the authors.

Figure 1
Figure 1. Moshi’s model architecture. It consists of the neu￾ral audio codec Mimi, which encodes speech waveforms into discrete audio tokens, and RQ-Transformer, which autoregres￾sively models sequences of text tokens and audio tokens. voices are recorded on separate channels. Furthermore, we aim to improve dialogue capabilities through data augmentation us￾ing 602 hours of synthetic spoken dialogue generated by multi￾stream … view at source ↗
Figure 2
Figure 2. Moshi’s Loss curve during pre-training on J-CHAT information. A token of layer 1 is called the semantic token, and those of layers 2-8 are called acoustic tokens. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Perplexity of speech generated by comparison models 4.2. Automatic Evaluation Results Following previous studies [1, 6], we evaluated the model’s flu￾ency by measuring the perplexity (PPL) of a language model6 on the ASR results of speech samples generated by each model. Results are shown in [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

36 extracted references · 27 canonical work pages

  1. [6]

    Conclusion In this study, we developed J-Moshi, a Japanese adaptation of Moshi, as a Japanese full-duplex spoken dialogue model. We conducted pre-training with 60,000 hours of Japanese speech data and fine-tuning with stereo spoken dialogue data and further attempted to improve performance using synthetic data generated through multi-stream TTS. In our ex...

    work page

  2. [1]

    In dialogue, full-duplex refers to having simultaneous bidirectional features such as speech overlaps and backchannels

    Introduction Full-duplex spoken dialogue systems are attracting attention as a way to achieve natural voice interaction [1, 2, 3]. In dialogue, full-duplex refers to having simultaneous bidirectional features such as speech overlaps and backchannels. Full-duplex spoken dialogue systems need to be researched for addressing the limi- tations of conventional...

    work page

  3. [2]

    Moshi This section explains the model architecture of Moshi [6] (Fig- ure 1), which consists of the neural speech codec Mimi and the large-scale speech language model RQ-Transformer. 2.1. Mimi Mimi is a neural speech codec [13] consisting of a SEANet [14] autoencoder and a residual vector quantizer [15]. The encoder discretizes 24,000Hz speech waveform da...

    work page Pith review arXiv 2025

  4. [3]

    J-Moshi

    Training in Japanese In this study, we adapt Moshi to Japanese spoken dialogue. This section explains the text vocabulary adaptation and two- stage training steps, consisting of pre-training and fine-tuning, that were implemented to adapt original Moshi 2 to Japanese. 2https://huggingface.co/kyutai/ moshiko-pytorch-bf16 We also explain data augmentation u...

    work page

  5. [4]

    This task involves generating a continuation from a few seconds of spoken dialogue prompt

    Experiments We adopted the prompted dialogue continuation task [1, 7], commonly used for evaluating full-duplex spoken dialogue models, to verify the spoken dialogue generation performance of J-Moshi and J-Moshi-ext through both automatic and human evaluations. This task involves generating a continuation from a few seconds of spoken dialogue prompt. We d...

    work page

  6. [5]

    Comparison of Japanese and English To analyze what kind of Japanese-specific behaviors J-Moshi acquired through our training, we compared it with the origi- nal Moshi. Specifically, we used four turn-taking statistics [1]: Inter-Pausal Units (IPU) (speech segments separated by at least 0.2 seconds of silence), Pause (silence between IPUs from the same spe...

    work page

  7. [7]

    We used the computational resources of the supercomputer “Flow” at the Information Technology Center, Nagoya University

    Acknowledgements This work was supported by JST Moonshot R&D, Grant num- ber JPMJMS2011. We used the computational resources of the supercomputer “Flow” at the Information Technology Center, Nagoya University

    work page

  8. [8]

    Omniflatten: An end-to-end gpt model for seamless voice conversation,

    Q. Zhang, L. Cheng, C. Deng, Q. Chen, W. Wang, S. Zheng, J. Liu, H. Yu, and C. Tan, “Omniflatten: An end-to-end gpt model for seamless voice conversation,” arXiv preprint arXiv:2410.17799, 2024

    arXiv 2024

Show all 36 references
  1. [9]

    Generative Spoken Dialogue Language Modeling,

    T. A. Nguyen, E. Kharitonov, J. Copet, Y . Adi, W.-N. Hsu, A. Elkahky, P. Tomasello, R. Algayres, B. Sagot, A. Mohamed, and E. Dupoux, “Generative Spoken Dialogue Language Modeling,” Transactions of the Association for Computational Linguistics, pp. 250–266, 2023

    2023

  2. [10]

    A Full-duplex Speech Dialogue Scheme Based On Large Language Model,

    P. Wang, S. Lu, Y . Tang, S. Yan, W. Xia, and Y . Xiong, “A Full-duplex Speech Dialogue Scheme Based On Large Language Model,” in Proceedings of the Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  3. [11]

    Language Model Can Listen While Speaking,

    Z. Ma, Y . Song, C. Du, J. Cong, Z. Chen, Y . Wang, Y . Wang, and X. Chen, “Language Model Can Listen While Speaking,” arXiv preprint arXiv:2408.02622, 2024

    2024 arXiv

  4. [12]

    SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilities,

    D. Zhang, S. Li, X. Zhang, J. Zhan, P. Wang, Y . Zhou, and X. Qiu, “SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilities,” in Findings of the Association for Computational Linguistics: EMNLP 2023 , 2023, pp. 15 757–15 773

    2023

  5. [13]

    LLaMA-Omni: Seamless Speech Interaction with Large Language Models,

    Q. Fang, S. Guo, Y . Zhou, Z. Ma, S. Zhang, and Y . Feng, “LLaMA-Omni: Seamless Speech Interaction with Large Language Models,” inProceedings of the Thirteenth International Conference on Learning Representations, 2025

    2025

  6. [14]

    Moshi: a speech-text foundation model for real-time dialogue,

    A. D ´efossez, L. Mazar´e, M. Orsini, A. Royer, P. P´erez, H. J´egou, E. Grave, and N. Zeghidour, “Moshi: a speech-text foundation model for real-time dialogue,” arXiv preprint arXiv:2410.00037, 2024

    2024 arXiv

  7. [15]

    Beyond Turn-Based Interfaces: Synchronous LLMs as Full- Duplex Dialogue Agents,

    B. Veluri, B. N. Peloquin, B. Yu, H. Gong, and S. Gollakota, “Beyond Turn-Based Interfaces: Synchronous LLMs as Full- Duplex Dialogue Agents,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , 2024, pp. 21 390–21 402

    2024

  8. [16]

    Robust Speech Recognition via Large-Scale Weak Supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. Mcleavey, and I. Sutskever, “Robust Speech Recognition via Large-Scale Weak Supervision,” in Proceedings of the 40th International Conference on Machine Learning, 2023, pp. 28 492–28 518

    2023

  9. [17]

    SALMONN-omni: A Codec- free LLM for Full-duplex Speech Understanding and Generation,

    W. Yu, S. Wang, X. Yang, X. Chen, X. Tian, J. Zhang, G. Sun, L. Lu, Y . Wang, and C. Zhang, “SALMONN-omni: A Codec- free LLM for Full-duplex Speech Understanding and Generation,” arXiv preprint arXiv:2411.18138, 2024

    2024 arXiv

  10. [18]

    J-CHAT: Japanese Large-scale Spoken Dialogue Corpus for Spoken Dialogue Language Modeling,

    W. Nakata, K. Seki, H. Yanaka, Y . Saito, S. Takamichi, and H. Saruwatari, “J-CHAT: Japanese Large-scale Spoken Dialogue Corpus for Spoken Dialogue Language Modeling,”arXiv preprint arXiv:2407.15828, 2024

    2024

  11. [19]

    Simultaneous talk—from the perspective of floor management of English and Japanese speakers,

    R. Hayashi, “Simultaneous talk—from the perspective of floor management of English and Japanese speakers,”World Englishes, vol. 7, no. 3, pp. 269–288, 1988

    1988

  12. [20]

    We also used the orig- inal 20-second audio (Ground-truth) as upper bound

    Additionally, to evaluate Mimi’s performance in Japanese spoken dialogue, we included simple re-synthesis of actual 20- second audio by Mimi (Re-synthesis). We also used the orig- inal 20-second audio (Ground-truth) as upper bound. As with the original Moshi, we tried three te...

  13. [21]

    Are you listening? cultural influences on the use of supportive verbal feedback in conversation,

    M. Stubbe, “Are you listening? cultural influences on the use of supportive verbal feedback in conversation,” Journal of Pragmatics, vol. 29, no. 3, pp. 257–289, 1998

    1998

  14. [22]

    High Fidelity Neural Audio Compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High Fidelity Neural Audio Compression,” Transactions on Machine Learning Research, 2023

    2023

  15. [23]

    SEANet: A multi-modal speech enhancement network,

    M. Tagliasacchi, Y . Li, K. Misiunas, and D. Roblek, “SEANet: A multi-modal speech enhancement network,” arXiv preprint arXiv:2009.02095, 2020

    2009 arXiv

  16. [24]

    Soundstream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “Soundstream: An end-to-end neural audio codec,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 30, pp. 495–507, 2021

    2021

  17. [25]

    SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing,

    T. Kudo and J. Richardson, “SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing,” in Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 2018, pp. 66–71

    2018

  18. [26]

    pyannote.audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,

    H. Bredin, “pyannote.audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,” inProceedings of the 24th INTER- SPEECH Conference, 2023, pp. 1983–1987

    2023

  19. [27]

    WhisperX: Time- Accurate Speech Transcription of Long-Form Audio,

    M. Bain, J. Huh, T. Han, and A. Zisserman, “WhisperX: Time- Accurate Speech Transcription of Long-Form Audio,” arXiv preprint arXiv:2303.00747, 2023

    2023 arXiv

  20. [28]

    ZeRO: memory optimizations toward training trillion parameter models,

    S. Rajbhandari, J. Rasley, O. Ruwase, and Y . He, “ZeRO: memory optimizations toward training trillion parameter models,” in Pro- ceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2020

    2020

  21. [29]

    Decoupled Weight Decay Regularization,

    I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regularization,” in Proceedings of the Sixth International Conference on Learning Representations, 2019

    2019

  22. [30]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, D. Bikel, L. Blecher, C. C. Ferrer, M. Chen, G. Cucurull, D. Es- iobu, J. Fernandes, J. Fu, W. Fu, B. Fuller, C. Gao, V . Goswami, N. Goyal, A. Hartshorn, S. ...

    2023 arXiv

  23. [31]

    CALLHOME Japanese Speech LDC96S37,

    A. Canavan and G. Zipperlen, “CALLHOME Japanese Speech LDC96S37,” Philadelphia: Linguistic Data Consortium, 1996

    1996

  24. [32]

    Corpus of Spontaneous Japanese: Its design and evaluation,

    K. Maekawa, “Corpus of Spontaneous Japanese: Its design and evaluation,” in Proceedings of ISCA & IEEE Workshop on Spon- taneous Speech Processing and Recognition, 2003

    2003

  25. [33]

    Travel agency task dialogue corpus: A multimodal dataset with age-diverse speakers,

    M. Inaba, Y . Chiba, Z. Qi, R. Higashinaka, K. Komatani, Y . Miyao, and T. Nagai, “Travel agency task dialogue corpus: A multimodal dataset with age-diverse speakers,” ACM Transac- tions on Asian and Low-Resource Language Information Process- ing, vol. 23, no. 9, pp. 1–23, 2024

    2024

  26. [34]

    Empirical Analysis of Training Strategies of Transformer-Based Japanese Chit-Chat Systems,

    H. Sugiyama, M. Mizukami, T. Arimoto, H. Narimatsu, Y . Chiba, H. Nakajima, and T. Meguro, “Empirical Analysis of Training Strategies of Transformer-Based Japanese Chit-Chat Systems,” in Proceedings of 2022 IEEE Spoken Language Technology Work- shop, 2023, pp. 685–691

    2022

  27. [35]

    Construction of a Japanese Daily Dialogue Corpus,

    R. Akama, Y . Isobe, J. Suzuki, and K. Inui, “Construction of a Japanese Daily Dialogue Corpus,” in Proceedings of the 29th Annual Conference of the Association for Natural Language Processing, 2023, pp. 108–113, (in Japanese)

    2023

  28. [36]

    RealPersonaChat: A realistic persona chat cor- pus with interlocutors’ own personalities,

    S. Yamashita, K. Inoue, A. Guo, S. Mochizuki, T. Kawahara, and R. Higashinaka, “RealPersonaChat: A realistic persona chat cor- pus with interlocutors’ own personalities,” in Proceedings of the 37th Pacific Asia Conference on Language, Information and Com- putation, 2023, pp. 852–861

    2023

Pith tools

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