Pith. sign in

REVIEW 3 major objections 4 minor 31 references

A two-stage dialogue framework lets a robot speak a context-aware preface early and deliver its main answer sooner.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 03:17 UTC pith:MY57E57N

load-bearing objection Missing initial-response duration undermines the claimed latency trade-off; the architecture is plausible but the central comparison is unverified. the 3 major comments →

arxiv 2607.23204 v1 pith:MY57E57N submitted 2026-07-25 cs.RO cs.CLcs.SD

Low-Latency Turn-Taking via Context-Aware Preface Generation in a Real-World Dialogue Robot

Yuki Okafuji , Koji Inoue , Yoshiki Ohira This is my paper
classification cs.RO cs.CLcs.SD
keywords turn-takingresponse latencydialogue robotincremental generationvoice activity projectionintent readiness detectionconversational fillershuman-robot interaction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that a dialogue robot can reduce the time users wait for a substantive answer by preparing a short, context-dependent preface as soon as the user's intent becomes predictable, and speaking it at a moment chosen by a turn-taking model. In a field experiment with a route-guidance robot in a shopping mall, prefaces cut the initial response latency roughly in half compared with no filler, and shortened the pause before the main answer from 0.74 s to 0.43 s compared with fixed fillers, at the cost of a slightly later first word (1.15 s vs 0.94 s). A sympathetic reader would care because it shows a concrete timing trade-off that can be engineered without waiting for fully end-to-end dialogue systems.

Core claim

The central claim is that decoupling prefatory-response preparation from speech onset — preparing a short content-aware preface once the user's intent is predictable, then delivering it when a voice-activity-projection model says the turn is yielding — yields faster access to the substantive answer than either no filler or a fixed filler. In a deployed shopping-mall robot, the contextual preface produced an initial-to-main gap of 0.43 s versus 0.74 s for a fixed filler, while keeping initial latency (1.15 s) far below the no-filler baseline (2.45 s) and only slightly above the fixed filler (0.94 s). The authors interpret the shorter gap as evidence of temporal overlap between main-response g

What carries the argument

The key mechanism is a two-stage incremental response architecture. An 'intent readiness detector' — a binary classifier trained on pseudo-labeled and human-annotated corpora — decides, from partial speech recognition, whether the user's intent is predictable enough to prepare a preface. A Voice Activity Projection (VAP) model, which predicts when the user will yield the turn, decides when to speak that preface. The preface (under 10 Japanese characters, non-committal) occupies the post-turn silence while the main LLM-based response is generated.

Load-bearing premise

The claim that contextual prefaces let the main response be generated during preface speech rests on the measured 'initial-to-main gap,' which is taken from the end of the initial response; since contextual prefaces are longer than fixed fillers, a later end point alone would shrink this gap even if the main-response generation started at the same absolute time.

What would settle it

Record the absolute time from user utterance end to the onset of the main response in both conditions. If the contextual-preface condition does not show a significantly earlier main-response onset than the fixed-filler condition (or does not approach the no-filler main-onset time), the shorter measured gap is an artifact of the metric rather than evidence of parallel generation. The same check could be done by re-analyzing the existing logs.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the framework holds in broader settings, LLM-based robots can deliver substantive answers about 0.9 s sooner than with no filler, without waiting for final recognition.
  • The separated preparation and speech-onset control makes the generation latency of the main response partially hidden behind speech, so the user perceives a shorter gap.
  • The intent-readiness detector firing before utterance end in 58.2% of cases means early preparation is often feasible in real interaction, not only in controlled settings.
  • The trade-off between initial latency and gap suggests designers can tune the trigger threshold to shift perceived responsiveness between first word and full answer.
  • Independent generation of preface and main response creates a risk of semantic/prosodic discontinuity, which the authors identify as the main bottleneck for quality.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the initial-to-main gap is measured from the end of the initial response, and contextual prefaces are longer than fixed fillers, part of the gap reduction may be mechanical: a later end point makes any fixed main-response onset time produce a smaller gap. The paper does not report absolute main-response onset times conditioned on the initial response, so the overlap interpretation is not f
  • One testable extension is to measure the absolute time from user utterance end to main-response onset, rather than the relative gap, which would separate true generation overlap from mere re-anchoring of the metric.
  • If the intent-readiness detector is made to trigger earlier, initial latency could approach the fixed-filler level while keeping the shorter gap; the breakdown analysis (20/251 prefaces) suggests the limit is the generation quality of short, grounded prefaces.
  • The framework is language- and domain-specific as presented; the same two-stage logic is likely to transfer to full-duplex systems, where the preface could be generated even before the user stops speaking.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a two-stage incremental response framework for LLM-based dialogue robots. An intent readiness detector triggers generation of a short contextual prefatory response when user intent becomes predictable, while a voice activity projection (VAP) model decides the precise moment to speak that preface; the main response is generated concurrently. A field experiment in a shopping mall compared three conditions (no-filler, fixed-filler, contextual-preface). Results show that both filler conditions significantly reduce initial response latency relative to no-filler, and that contextual-preface has significantly longer initial latency than fixed-filler but a significantly shorter initial-to-main gap. Subjective ratings show no significant differences. The paper interprets this as a timing trade-off: the contextual preface masks turn-onset latency and also shortens the wait for the substantive answer, at the cost of slightly later first speech.

Significance. If the timing effects are robust, the work offers a practical architectural pattern for reducing perceived response delay in cascaded ASR-LLM-TTS robots without relying on generic fillers. The field deployment is a real strength, as is the honest treatment of null subjective results and the explicit acknowledgment of limitations. The paper also provides a breakdown analysis of prefatory-response failures, which is valuable for future work. However, the central quantitative claim that contextual prefaces deliver the substantive answer sooner is not directly supported by the reported metrics, because the duration of the initial response is omitted. The paper's significance therefore hinges on whether the missing timing information actually confirms the claimed trade-off.

major comments (3)
  1. [Section 3.2, Table 1] The time from user utterance end to main-response onset is M = L + D + G, where L is initial latency, D is the duration of the initial response, and G is the initial-to-main gap. The paper reports L and G but not D. Since contextual prefaces are constrained to be at most 10 Japanese characters while fixed fillers are 'yeah'/'I see' (Section 2.1), D_ctx likely exceeds D_fixed. The reported advantage in L+G is only about 0.10 s (1.58 vs 1.68 s); if D_ctx exceeds D_fixed by more than this, the substantive answer actually arrives later in the contextual-preface condition. To support the Discussion's claim that the method shortened the delay before the substantive answer (Section 4), the authors must report D and M, or at least the absolute main-response onset time conditioned on the initial response.
  2. [Section 3.2, metric definition] The initial-to-main gap is measured from the end of the initial response. This makes the comparison between fixed-filler and contextual-preface structurally sensitive to the duration of the initial response: a longer initial response moves the measurement window later, mechanically reducing the gap even if main-response onset timing is identical. The paper does not control for initial-response duration or restrict the gap analysis to responses with matched initial-response lengths. The authors should either analyze the absolute main-response onset time (as in the previous comment) or rerun the gap comparison including initial-response duration as a covariate/per-utterance matched analysis.
  3. [Section 3.3, statistical tests] The Kruskal-Wallis and Mann-Whitney tests treat each response instance as an independent sample. However, the data contain multiple response instances from the same interaction (47-75 interactions, with 188-251 response instances per condition). Responses within an interaction are likely correlated (same user, same context, same system state). This pseudoreplication can inflate the effective sample size and render the reported p-values (e.g., p < .001 for the gap difference) overly optimistic. The authors should fit a mixed-effects model with interaction as a random intercept, or at least perform a sensitivity analysis using one aggregated value per interaction (e.g., per-interaction medians) and report whether the qualitative conclusions hold.
minor comments (4)
  1. [Section 3.1] Typo: 'ano-filler condition' should be 'a no-filler condition'.
  2. [Figure 1] The figure caption appears truncated or confusing: the text '('McDonald's would be nice.')' is separated across lines, and the speech-bubble labels are not fully legible in the reproduced version. Please clean up the caption and ensure all callouts in the figure appear in the text or are explained.
  3. [Section 3.2] The 'Trigger position' metric is defined as a character-level progress ratio, but it is not described how the character counts are computed (e.g., including punctuation, hesitations). Clarify this definition.
  4. [Section 3.3] The sentence 'both fixed-filler and contextual-preface were significantly faster than no-filler' lacks the actual p-values for those pairwise comparisons; only the fixed-filler vs contextual-preface comparison is reported (p=.027). Please report all pairwise results or state that they were significant at the Bonferroni-corrected threshold.

Circularity Check

0 steps flagged

No significant circularity: the latency results are measured field observations, not predictions fitted to the claims.

full rationale

This paper is an empirical field-evaluation paper, not a formal derivation, so the main circularity failure modes (self-definitional prediction, fitted input renamed as prediction, uniqueness imported from authors, ansatz smuggled via citation) do not apply. The central results—initial response latency and initial-to-main gap—are reported as measured statistics from a deployed robot (Table 1: no-filler 2.45 s, fixed-filler 0.94 s/0.74 s, contextual-preface 1.15 s/0.43 s), not as quantities derived from a fitted parameter or from a prior theorem. The intent-readiness threshold of 0.35 is described as 'empirically set,' but it is a deployment threshold for triggering preface generation, not a parameter tuned to produce the latency comparisons, and the comparisons are robust under any fixed threshold. The paper cites the authors' own prior VAP-based dialogue system [10] as the platform that is extended, and uses VAP [9] for turn-end prediction; however, all three experimental conditions share the same VAP-based turn-taking mechanism, so this self-citation is component reuse rather than load-bearing evidence for the paper's comparative claim. A reviewer concern is that the initial-to-main gap is defined as 'the time from the end of the initial response to the beginning of the main response,' which is only a component of the absolute time from user utterance end to main-response onset (L + D + G). Since contextual prefaces are constrained to 'no longer than 10 Japanese characters' whereas fixed fillers are 'yeah' or 'I see,' the shorter observed gap could be partly mechanical if the contextual initial response is longer; the paper does not report initial-response durations or absolute main-onset times. This is a real evidentiary/validity limitation in the interpretation that contextual prefaces 'shortened the gap before the substantive response,' but it is not circularity: the gap statistic is measured, not defined in terms of the conclusion, and the paper's own stated claim is a timing trade-off rather than a fitted 'prediction.' The paper also explicitly acknowledges its limitations (single Japanese shopping-mall setting, dependence on specific ASR/LLM/TTS components, exploratory questionnaire sample). Accordingly, no step in this paper reduces by construction to its inputs, and the appropriate score is at the no-substantial-circularity end of the scale.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

No new physical or theoretical entities are introduced. The framework uses existing LLM, STT, and VAP components as building blocks. The free parameters are deployment choices rather than scientific constants. The main axioms are domain assumptions about the validity of the intent-readiness labels, the reliability of the VAP model in the field, and the statistical treatment of repeated observations.

free parameters (2)
  • Intent readiness decision threshold = 0.35
    Empirically set in Section 2.2; controls how early prefatory generation is triggered, which directly affects initial latency and initial-to-main gap; no sensitivity analysis is provided.
  • Prefatory response length limit = ≤10 Japanese characters
    Hand-chosen design constraint in Section 2.1; affects the duration of the initial response and therefore the definition of the initial-to-main gap.
axioms (5)
  • domain assumption Human-annotated labels for intent readiness are valid ground truth
    The detector is fine-tuned on 30,740 human-annotated instances (Section 2.2); the field evaluation assumes this readiness signal is meaningful and well-calibrated.
  • domain assumption The VAP turn-taking model from prior work [10] gives accurate turn-end predictions in this shopping-mall environment
    Delivery timing relies on a locally running VAP model with fallback to final STT (Section 2.3); differences in VAP/STT trigger rates affect the measured latencies.
  • domain assumption Each response instance can be treated as an independent observation for hypothesis testing
    Section 3.3 uses Kruskal-Wallis and Mann-Whitney tests on response instances; multiple responses come from the same interaction, violating independence and potentially inflating significance.
  • domain assumption System log timestamps accurately capture user utterance end and response onset, and the exclusion of STT delay is consistent across conditions
    The latency metric is defined in Section 3.2 as time from end of user utterance to first response excluding the delay before the STT result; if the definition applies differently per condition, comparisons are biased.
  • standard math Kruskal-Wallis and Mann-Whitney U tests are appropriate for skewed latency distributions
    Standard non-parametric procedures used because latency values are skewed; this is uncontroversial except for the independence issue already noted.

pith-pipeline@v1.3.0-alltime-deepseek · 7701 in / 12206 out tokens · 113118 ms · 2026-08-01T03:17:21.298012+00:00 · methodology

0 comments
read the original abstract

Large language model (LLM)-based dialogue systems suffer response delays because generation begins only after final speech recognition. While fixed fillers are a workaround, they become unnatural over time. We propose a two-stage incremental framework that decouples prefatory-response preparation from speech onset. Once user intent becomes predictable, an intent readiness detector triggers LLM-based generation of a short prefatory response. Concurrently, a voice activity projection (VAP) model determines when to deliver it. Through a field experiment with a route-guidance robot in a shopping mall, we evaluated three conditions: no-filler, fixed-filler, and contextual-preface. Both fixed-filler and contextual-preface significantly reduced initial response latency relative to no-filler. Relative to fixed-filler, contextual-preface had significantly longer initial response latency but a significantly shorter initial-to-main gap. Exploratory ratings showed no significant differences. These results indicate a timing trade-off.

Figures

Figures reproduced from arXiv: 2607.23204 by Koji Inoue, Yoshiki Ohira, Yuki Okafuji.

Figure 1
Figure 1. Figure 1: Proposed two-stage architecture for incremental response generation. The system prepares a prefatory response [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example of users interacting with robots. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of initial response latency [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of initial-to-main gap Separating turn-entry sources into VAP and STT further clar￾ified the source of the latency differences. VAP-triggered onsets accounted for 73.2%, 78.7%, and 77.7% of responses in the no-filler, fixed-filler, and contextual-preface conditions, respectively. The cor￾responding mean initial response latencies for VAP-triggered onsets were 2.20, 0.68, and 0.92 s, whereas th… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 3 canonical work pages

  1. [1]

    Halim-Antoine Boukaram, Micheline Ziadee, and Majd F Sakr. 2021. Mitigating the Effects of Delayed Virtual Agent Response Time Using Conversational Fillers. InProceedings of the 9th International Conference on Human-Agent Interaction (Virtual Event, Japan)(HAI ’21). Association for Computing Machinery, New York, NY, USA, 130–138. doi:10.1145/3472307.3484181

    arXiv 2021

  2. [2]

    Yuya Chiba and Ryuichiro Higashinaka. 2025. Investigating the Impact of Incre- mental Processing and Voice Activity Projection on Spoken Dialogue Systems. InProceedings of the 31st International Conference on Computational Linguistics, Owen Rambow, Leo Wanner, Marianna Apidianaki, Hend Al-Khalifa, Barbara Di Eugenio, and Steven Schockaert (Eds.). Associat...

    2025

  3. [3]

    Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, and Neil Zeghidour. 2024. Moshi: a speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037(2024). https://doi.org/10.48550/arXiv.2410.00037

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.00037 2024

  4. [4]

    David DeVault, Kenji Sagae, and David Traum. 2009. Can I finish? Learning when to respond to incremental interpretation results in interactive dialogue. In Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL). 11–20

    2009

  5. [5]

    Erik Ekstedt and Gabriel Skantze. 2020. TurnGPT: a Transformer-based Language Model for Predicting Turn-taking in Spoken Dialog. InFindings of the Association for Computational Linguistics: EMNLP 2020, Trevor Cohn, Yulan He, and Yang Liu (Eds.). Association for Computational Linguistics, Online, 2981–2990. doi:10. 18653/v1/2020.findings-emnlp.268

    2020

  6. [6]

    Erik Ekstedt and Gabriel Skantze. 2022. Voice Activity Projection: Self-supervised Learning of Turn-taking Events. InInterspeech 2022. 5190–5194. doi:10.21437/ Interspeech.2022-10955

    2022

  7. [7]

    Iván Esteban-Lozano, Álvaro Castro-González, and Paloma Martínez. 2024. Us- ing a LLM-Based Conversational Agent in the Social Robot Mini. InArtificial Intelligence in HCI, Helmut Degen and Stavroula Ntoa (Eds.). Springer Nature Switzerland, Cham, 15–26. https://doi.org/10.1007/978-3-031-60615-1_2

    work page doi:10.1007/978-3-031-60615-1_2 2024

  8. [8]

    Ryuichiro Higashinaka, Masahiro Araki, Hiroshi Tsukahara, and Masahiro Mizukami. 2021. Integrated taxonomy of errors in chat-oriented dialogue sys- tems. InAnnual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL). 89–98

    2021

  9. [9]

    Koji Inoue, Bing’er Jiang, Erik Ekstedt, Tatsuya Kawahara, and Gabriel Skantze

  10. [10]

    Koji Inoue, Yuki Okafuji, Jun Baba, Yoshiki Ohira, Katsuya Hyodo, and Tatsuya Kawahara. 2025. A Noise-Robust Turn-Taking System for Real-World Dialogue Robots: A Field Experiment. In2025 IEEE/RSJ International Conference on Intelli- gent Robots and Systems (IROS). 874–879. doi:10.1109/IROS60139.2025.11246533

    arXiv 2025

  11. [11]

    Yuin Jeong, Juho Lee, and Younah Kang. 2019. Exploring Effects of Conversational Fillers on User Perception of Conversational Agents. InExtended Abstracts of the 2019 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, Article LBW2715, 6 pages. doi:10.1145/3290607.3312913

    arXiv 2019

  12. [12]

    Dahyun Kang, Changjoo Nam, and Sonya S. Kwak. 2024. Robot Feedback Design for Response Delay.International Journal of Social Robotics16 (2024), 341–361. doi:10.1007/s12369-023-01068-z

    work page doi:10.1007/s12369-023-01068-z 2024

  13. [13]

    Divesh Lala, Shizuka Nakamura, and Tatsuya Kawahara. 2019. Analysis of Effect and Timing of Fillers in Natural Turn-Taking. InINTERSPEECH. 4175–4179

    2019

  14. [14]

    Levinson and Francisco Torreira

    Stephen C. Levinson and Francisco Torreira. 2015. Timing in turn-taking and its implications for processing models of language.Frontiers in Psychology6 (2015),

    2015

  15. [15]

    McMahan, and Joseph J

    Mykola Maslych, Mohammadreza Katebi, Christopher Lee, Yahya Hmaiti, Amirpouya Ghasemaghaei, Christian Pumarada, Janneese Palmer, Esteban Segarra Martinez, Marco Emporio, Warren Snipes, Ryan P. McMahan, and Joseph J. LaViola Jr. 2025. Mitigating Response Delays in Free-Form Conver- sations with LLM-powered Intelligent Virtual Agents. InACM Conference on Co...

    arXiv 2025

  16. [16]

    Kiyotada Mori, Seiya Kawano, Angel García Contreras, and Koichiro Yoshino

  17. [17]

    Ryosuke Nakanishi, Koji Inoue, Shizuka Nakamura, Katsuya Takanashi, and Tatsuya Kawahara. 2018. Generating Fillers Based on Dialog Act Pairs for Smooth Turn-Taking by Humanoid Robot. InInternational Workshop on Spoken Dialogue Systems Technology. https://api.semanticscholar.org/CorpusID:52572791

    2018

  18. [18]

    Masaya Ohagi, Tomoya Mizumoto, and Katsumasa Yoshikawa. 2024. Investigation of look-ahead techniques to improve response time in spoken dialogue system. InINTERSPEECH. 3580–3584

    2024

  19. [19]

    Matthew Roddy, Gabriel Skantze, and Naomi Harte. 2018. Investigating Speech Features for Continuous Turn-Taking Prediction Using LSTMs. InInterspeech

    2018

  20. [20]

    Rajarshi Roy, Jonathan Raiman, Sang-gil Lee, Teodor-Dumitru Ene, Robert Kirby, Sungwon Kim, Jaehyeon Kim, and Bryan Catanzaro. 2026. PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models.arXiv preprint arXiv:2602.06053(2026). https://doi.org/10.48550/arXiv.2602.06053

    doi:10.48550/arxiv.2602.06053 2026

  21. [21]

    David Schlangen and Gabriel Skantze. 2009. A General, Abstract Model of Incre- mental Dialogue Processing. InProceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009). Association for Computational Linguistics, Athens, Greece, 710–718. https://aclanthology.org/E09-1081/

    2009

  22. [22]

    Toshiyuki Shiwa, Takayuki Kanda, Michita Imai, Hiroshi Ishiguro, and Norihiro Hagita. 2009. How Quickly Should a Communication Robot Respond? Delaying Strategies and Habituation Effects.International Journal of Social Robotics1, 2 (2009), 141–155. doi:10.1007/s12369-009-0012-8

    doi:10.1007/s12369-009-0012-8 2009

  23. [23]

    Gabriel Skantze. 2021. Turn-taking in conversational systems and human-robot interaction: A review.Computer Speech & Language67 (2021), 101178. doi:10. 1016/j.csl.2020.101178

    arXiv 2021

  24. [24]

    Gabriel Skantze and Bahar Irfan. 2025. Applying General Turn-taking Models to Conversational Human-Robot Interaction. InProceedings of the 2025 ACM/IEEE International Conference on Human-Robot Interaction(Melbourne, Australia)(HRI ’25). IEEE Press, 859–868. https://dl.acm.org/doi/10.5555/3721488.3721593

    arXiv 2025

  25. [25]

    Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hall- ström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli. 2024. Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Mem- ory Efficient, and Long Context Finetuning and Inf...

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.13663 2024

  26. [26]

    Jianing Yang, Yusuke Fujita, and Yui Sudo. 2026. DuplexCascade: Full-Duplex Speech-to-Speech Dialogue with VAD-Free Cascaded ASR-LLM-TTS Pipeline and Micro-Turn Optimization.arXiv preprint arXiv:2603.09180(2026). https: //doi.org/10.48550/arXiv.2603.09180

    doi:10.48550/arxiv.2603.09180 2026

  27. [27]

    Ceng Zhang, Junxin Chen, Jiatong Li, Yanhong Peng, and Zebing Mao. 2023. Large language models for human–robot interaction: A review.Biomimetic Intelligence and Robotics3, 4 (2023), 100131. doi:10.1016/j.birob.2023.100131

    arXiv 2023

  28. [731]

    doi:10.3389/fpsyg.2015.00731

    arXiv 2015

  29. [2018]

    doi:10.21437/Interspeech.2018-2124

    586–590. doi:10.21437/Interspeech.2018-2124

    work page doi:10.21437/interspeech.2018-2124 2018

  30. [2024]

    Accepted to the 14th In- ternational Workshop on Spoken Dialogue Systems Technology (IWSDS 2024)

    Real-time and Continuous Turn-taking Prediction Using Voice Activity Projection.arXiv preprint arXiv:2401.04868(2024). Accepted to the 14th In- ternational Workshop on Spoken Dialogue Systems Technology (IWSDS 2024). doi:10.48550/arXiv.2401.04868

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.04868 2024

  31. [2025]

    InINTERSPEECH

    Dialogue Response Prefetching Based on Semantic Similarity and Prediction Confidence of Language Model. InINTERSPEECH. 3045–3049