acceptodds
Under review as a conference paper at ICLR 2027

DuplexTTS: A Scalable Framework for Full-Duplex Conversational Speech Synthesis and Evaluation

Abstract

The scarcity of full-duplex conversational speech data limits the development of natural spoken interaction models. Conventional text-to-speech systems primarily generate isolated utterances or alternating turns, leaving the joint generation of conversational content and interaction timing underexplored. Motivated by advances in native full-duplex speech generation, we introduce DuplexTTS, a scalable framework for synthesizing and evaluating full-duplex conversational speech. DuplexTTS supports two synthesis routes: native autonomous generation, where a single speech model generates two role-conditioned streams through shared acoustic feedback, and rule-based orchestration, where scripted utterances are synthesized and arranged into turn exchanges, backchannels, interruptions, and side talk. A unified pipeline connects synthesis with dialogue-level quality screening, additional source-utterance checks for rule synthesis, interaction analysis, and downstream evaluation. Using this framework, we construct over 1,800 hours of screened Chinese and English dialogue. Human listeners rate the generated dialogues favorably for naturalness and meaningfulness. Training Moshi on the two corpora yields different conversational behaviors. The mixed-data model achieves 16.60 percentage points higher macro expected-behavior accuracy on Full-Duplex-Bench than the original model. These results establish DuplexTTS as a practical framework for expanding full-duplex training data and studying how synthesis and curation shape spoken interaction. Audio demos are available at https://duplextts.github.io.

Primary area:
datasets and benchmarks
Keywords:
full-duplex dialogue, conversational speech synthesis, synthetic speech data, spoken dialogue modeling, automated quality assessment

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_u4WhOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents DuplexTTS, a framework for synthesizing and screening full-duplex conversational speech through native generation and rule-based composition. The paper construct a bilingual speech corpus and assess its quality and utility through human evaluation and downstream full-duplex model training.

Critical strengths
  1. Complementary synthesis routes with systematic quality control. Native and rule-based synthesis produce corpora with complementary downstream strengths in interruption response and continuation through non-interrupting speech. Final-audio screening and additional source-utterance checks strengthen this mixed-data pipeline by detecting missing or near-silent utterances that can escape dialogue-level checks.

  2. Evidence of downstream utility supported by multi-level evaluation. The authors assess the generated data through interaction statistics and human listening, and further evaluate its utility through Moshi adaptation. Mixed-data training increases macro expected-behavior accuracy on Full-Duplex-Bench from 26.43% to 43.03%, providing evidence of practical value beyond perceptual quality alone.

Critical weaknesses
  1. Insufficient comparison with existing data-construction approaches. The paper lacks both a structured comparison of its capabilities against related pipelines and a downstream comparison with Moshi fine-tuned on a suitable alternative full-duplex corpus. Improvement over the unadapted Moshi checkpoint alone cannot distinguish the value of this particular framework from the benefit of additional task-relevant training. Without these comparisons, the methodological differentiation and practical advantage of the proposed synthesis and curation choices remain insufficiently established.

  2. The downstream comparisons confound data composition with data quantity and training budget. The mixed condition uses 1,425.61 hours of training data and is evaluated at update 9,100, whereas rule-only uses 167.97 hours and is evaluated at update 700. Consequently, the stronger mixed-data result cannot establish whether combining the two synthesis routes provides benefits beyond using more data and optimization. Matched-budget comparisons are needed to substantiate the incremental value of the mixed-data design.

Questions for authors
  1. Would it be possible to add a table comparing DuplexTTS with Lychee-FD[1], DuplexGen[2], ConversationalVoice[3], DuplexChat[4], and SmoothConv & DuplexConv[5] in terms of data size, data format, construction methods, and quality checks?

  2. How does Moshi trained on DuplexTTS compare with the same model trained on at least one suitable existing dataset, using the same data size and training budget? Alternatively, how does it compare with a suitable released model on the same benchmark?

  3. What are the numbers and proportions of normal turns, backchannels, interruptions, and side talk in the rule-based training set?

  4. Does mixed-data training still outperform rule-only training after applying the interruption-timing correction from Appendix B.3, when data size and training budget are matched?

[1] Liu Z, Zhang X, Li Y, et al. Hierarchical Acoustic-Semantic Modeling: Modality Separation and Semantic Coherence for Full-Duplex SLMs[C]//Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026: 9264-9280. [2] Kim T, Kim K, Woo S H, et al. DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues[J]. arXiv preprint arXiv:2607.26178, 2026. [3] He R Y, Cao B, Xu C, et al. ConversationalVoice: Full-Duplex Speech Data from Real Conversations through Source-Faithful Reconstruction and Conversation-Grounded Expansion[J]. arXiv preprint arXiv:2609.08147, 2026. [4] Nakata W, Saito Y, Saruwatari H. DuplexChat: Constructing Speaker-Separated Full-Duplex Dialogue Speech at Scale for Spoken Dialogue Language Modeling[J]. arXiv preprint arXiv:2607.04941, 2026. [5] Wang C, Shao M, He C, et al. SmoothConv and DuplexConv: Complementary Mandarin Multi-Party Conversational Speech Corpora for Speech Interaction[J]. arXiv preprint arXiv:2610.11150, 2026.

7 comments
AuthorsOpenReview ↗

Response to Reviewer u4Wh

Dear Reviewer u4Wh,

Thank you for the detailed review, the five references, and the two concrete experimental requests — both now carried out.

W1 & Q1 (structured comparison). The revision adds the requested table as Table 1 (Section 2, p. 2), organized along the dimensions the review asked for: data size (each paper's own counting unit), data format, construction method, and quality checks, plus downstream training evaluation, for all five suggested pipelines and DuplexTTS. Every cell is verified against the primary papers and official releases (notes in Appendix E.6); unreported properties are marked n.r. The comparison shows where DuplexTTS differs: existing pipelines render scripted exchanges through TTS cascades or source real conversations, while DuplexTTS combines native generation, where a single speech model produces both interacting channels, with rule-based composition over verified source utterances. We discuss the downstream comparison under Q2.

W2 & Q4 (training budget, and mixing after the timing correction). We agree with this characterization of the original downstream comparison; the global response reports the matched-budget experiment, and the revision scopes the full-pool mixed result as recipe-level. The same experiment also addresses Q4. Timing correction reduces the rule-only macro score by 5.11 pp but has little effect on mixed training (−0.14 pp). After correction, mixed training still has a lower macro score than rule-only training (−3.73 pp [−7.2, −0.4]), but performs better on interruption response (+24.50 pp [+19.2, +30.0]). (Table 5, p. 9; contrasts in Table 20, Appendix E.1.)

Q2 (released-model comparison). Following the review's alternative, we evaluated the released Lychee-FD model on the identical 498-case protocol: it reaches 27.03 macro, versus 24.42 for the original checkpoint and 33.81–42.65 for every matched DuplexTTS-trained arm. Lychee-FD differs in architecture and training history, and its frequent unknown judgments (e.g., 68 of 98 backchannel cases) remain in the denominator. We also considered training the same backbone on an external corpus at equal budget: the DuplexGen release lacks word timestamps and would need a new alignment stage, which did not fit the discussion window. (Table 5, p. 9; Appendix E.2.)

Q3 (event composition). The revision reports counts over the actual training manifests: for the original rule-only subset, normal 48.8%, backchannel 19.5%, interruption 17.1%, side talk 14.6% of 66,951 unique events, with all six rule subsets in Appendix E.5 (Table 21); per-window occurrence counts shift each share by at most 2.4 points.

We hope these additions provide a useful basis for discussion, and we would be glad to provide further detail during the discussion period.

Best regards,

The Authors

AuthorsOpenReview ↗

Supplementary tables for Q1–Q4 (inline)

Dear Reviewer u4Wh,

For the reviewer's convenience, we provide the tables supporting our earlier reply directly in this thread, each followed by a brief summary. All values match the revised PDF.

Q1 — the requested comparison (Table 1 of the revision). Sizes use each paper's own counting unit; n.r. = not reported. Columns: separate per-speaker channels, word timestamps, controlled event placement, source-utterance checks, dialogue screening, downstream full-duplex training evaluation. Source-verified notes for every cell are in Appendix E.6.

PipelineConstructionSize (own unit)Sep.TSEventSrc.ScreenFD eval.
Lychee-FD [1]agents + TTS~140K dial.✔️n.r.✔️–✔️✔️
DuplexGen [2]TTS + sched. placement1,623 h✔️–✔️–✔️✔️
ConversationalVoice [3]real-audio reconstr.n.r.✔️✔️–✔️✔️–
DuplexChat [4]podcast reconstruction415K h (URLs only)✔️–––✔️n.r.
SmoothConv & DuplexConv [5]spontaneous rec.2,100 h✔️✔️––✔️–
DuplexTTS (ours)native gen. + rule comp.1,835 h✔️✔️✔️✔️✔️✔️

Existing pipelines render scripted exchanges through TTS cascades or source real conversations, while DuplexTTS (ours) combines native generation, where a single speech model produces both interacting channels, with rule-based composition over verified source utterances.

Q2 — released-model comparison, per task (reference rows of Table 5).

TaskLychee-FD (released)Original Moshi (re-eval.)Matched DuplexTTS arms (range)
Interruption55.5058.5030.67–59.00
Backchannel30.618.1619.73–29.25
Talking-to-other8.0021.0030.33–53.67
Background14.0010.0024.00–49.00
Macro27.0324.4233.81–42.65

Lychee-FD[1] improves over the original checkpoint mainly through backchannel continuation; every matched DuplexTTS arm reaches a higher macro. As noted earlier, Lychee-FD differs in architecture and training history, and its unknown judgments remain in the denominator.

Q3 — event composition of all rule training subsets (Table 21 of the revision). Unique retained source events; count and share of each row's total. Per-window occurrence counts shift each share by at most 2.4 points.

Training subsetNormalBackchannelInterruptionSide talkTotal
Original rule-only32,691 (48.8%)13,083 (19.5%)11,431 (17.1%)9,746 (14.6%)66,951
Rule part of original mixed97,364 (48.6%)39,348 (19.6%)34,371 (17.2%)29,338 (14.6%)200,421
Matched rule-only, original22,622 (47.7%)10,740 (22.7%)7,134 (15.1%)6,886 (14.5%)47,382
Matched rule-only, corrected22,711 (47.7%)10,834 (22.8%)7,162 (15.0%)6,899 (14.5%)47,606
Rule part of matched mixed, original11,422 (47.9%)5,400 (22.6%)3,570 (15.0%)3,458 (14.5%)23,850
Rule part of matched mixed, corrected11,460 (47.9%)5,441 (22.7%)3,586 (15.0%)3,462 (14.5%)23,949

The four event types keep stable shares across all rule subsets, so the matched arms train on the same event mixture as the original rule corpus.

AuthorsOpenReview ↗

Supplementary tables for Q1–Q4 (inline) (2)

Q4 (and W2) — the five prespecified contrasts (Table 20 of the revision). Percentage points; 95% case-bootstrap intervals; three paired seeds. The interaction row subtracts the two correction rows: a positive value means the timing correction costs rule-only training more than mixed training.

ContrastEqual-task macroInterruption
Mixed − Rule (original timing)−8.70 [−12.5, −4.8]+20.33 [+14.3, +26.2]
Mixed − Rule (corrected timing)−3.73 [−7.2, −0.4]+24.50 [+19.2, +30.0]
Correction effect on Rule-only−5.11 [−8.0, −2.2]−8.00 [−12.7, −3.3]
Correction effect on Mixed−0.14 [−2.8, +2.4]−3.83 [−8.3, +0.7]
Correction × composition+4.97 [+1.1, +8.7]+4.17 [−2.0, +10.3]

At matched budget, the macro advantage of mixing does not hold under either timing condition, while its interruption advantage persists and grows after the correction, which costs rule-only training more than mixed training.

We hope these tables are helpful, and we would be glad to provide further detail during the discussion period.

Best regards, The Authors

[1] Liu Z, Zhang X, Li Y, et al. Hierarchical Acoustic-Semantic Modeling: Modality Separation and Semantic Coherence for Full-Duplex SLMs[C]//Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026: 9264-9280. [2] Kim T, Kim K, Woo S H, et al. DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues[J]. arXiv preprint arXiv:2607.26178, 2026. [3] He R Y, Cao B, Xu C, et al. ConversationalVoice: Full-Duplex Speech Data from Real Conversations through Source-Faithful Reconstruction and Conversation-Grounded Expansion[J]. arXiv preprint arXiv:2609.08147, 2026. [4] Nakata W, Saito Y, Saruwatari H. DuplexChat: Constructing Speaker-Separated Full-Duplex Dialogue Speech at Scale for Spoken Dialogue Language Modeling[J]. arXiv preprint arXiv:2607.04941, 2026. [5] Wang C, Shao M, He C, et al. SmoothConv and DuplexConv: Complementary Mandarin Multi-Party Conversational Speech Corpora for Speech Interaction[J]. arXiv preprint arXiv:2610.11150, 2026.

Reviewer_crWPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

DuplexTTS proposes a framework for scaling full-duplex dialogue data through native and rule-based synthesis. The native route shares PersonaPlex weights across roles while maintaining separate states and using acoustic feedback, with backbone adaptation followed by acoustic distillation. The more conventional rule-based route synthesizes utterances independently and arranges interactions through temporal alignment and event scheduling. The pipeline constructs 1,835.3 hours of data using audio screening and supervision alignment, evaluated through corpus analysis, human listening, and Moshi fine-tuning. English mixed-data training improves Full-Duplex-Bench macro expected-behavior accuracy from 26.43% to 43.03%.

Critical strengths
  1. Downstream training utility. Mixed-data fine-tuning improves all three continuation tasks over the original Moshi and reduces non-interruption responses from 40/298 to 13/298. These results demonstrate the synthetic data’s training utility for improving interaction behavior on the evaluated benchmark.
  2. Utterance completeness and supervision alignment. Source-level checks flag 12,977 additional dialogues containing missing or near-silent clips. The English rule export also masks words removed by interruption cutoffs or affected by ambiguous same-channel mixtures. These procedures separately address utterance omissions and audio–label mismatches.
Critical weaknesses

The route comparison does not control for data volume and training budget. In Table 13, Native, Rule, and Mixed use 614.13, 167.97, and 1,425.61 hours, with evaluated checkpoints selected at updates 2,000, 700, and 9,100, respectively. Since synthesis route, data volume, and optimization vary together, the results support recipe-level utility but do not isolate the benefit of adding native data under matched budgets.

Questions for authors
  1. Matched-budget mixing. Has rule-only training been compared with mixed-data training under matched training-audio volume and optimization schedules?
  2. Post-interruption supervision. How are B’s target utterances constructed after A interrupts in the English rule-based data? Appendix B.2 describes B’s cutoff, and D.1 uses B as the training output, whereas benchmark success (C_RESPOND) requires addressing the interruption’s content rather than merely stopping. A training transcript and B-side targets spanning an interruption would clarify the connection between the data and evaluation.
  3. Human assessment of adaptation. Has human evaluation assessed the effect of native adaptation itself? In particular, is there an independent comparison between the original PersonaPlex and the adapted native generator under matched conditions?
5 comments
AuthorsOpenReview ↗

Response to Reviewer crWP

Dear Reviewer crWP,

Thank you for recognizing the downstream training utility, the source-utterance completeness checks, and the masking design, and for three questions that each improved the revision.

W1 & Q1 (matched-budget mixing). We agree that the original comparison does not isolate the benefit of adding native data; details can be found in the global response. Under matched budgets, rule-only training has the higher macro while mixed training responds to interruptions more often. The revision adds this controlled comparison and presents the original result as recipe-level evidence. (Section 6, Table 5, p. 9.)

Q2 (B-side supervision after A interrupts). The original Appendix B.2 did not explain how B’s reply is supervised after the cutoff. The revision now walks through the construction end to end on one real training window: B's cutoff words are masked with zero loss weight, the frames after the cutoff carry padding targets (the model learns to stop), and B's next scripted utterance — written as a reply to the interruption — enters the window as unmasked training targets (the model learns to respond). In the audited example, A interrupts with "Present tense might be too jarring right after the prologue," and B's supervised reply is "That's fair. Stick with past tense then, but tighten the verbs." A boundary example also shows the limitation: when the reply falls outside the 60-second window, that view contributes no response tokens. (Timeline in Figure 8, Appendix E.4, p. 27; pointer in Section 3.3.)

Q3 (independent human assessment of adaptation). We ran the requested study: three independent listeners blindly scored matched 40-clip packets from the original PersonaPlex and the final adapted generator (same scenes, roles, voices, seeds, and sampling per pair).

Scale (1–5)OriginalAdaptedPaired diff95% CI
Interaction naturalness (N-MOS)3.484.17+0.69[+0.39, +0.99]
Content meaningfulness (M-MOS)4.094.41+0.35[+0.11, +0.60]
Setting and role adherence3.954.32+0.38[+0.16, +0.60]
Channel A audio quality4.724.69−0.03[−0.14, +0.08]
Channel B audio quality4.774.69−0.07[−0.18, +0.03]

The results shows that all three dialogue-level scales improve; audio quality is nearly unchanged. (Table 3 now sits beside the adaptation scores of Table 2, Section 3.1, p. 5; protocol in Appendix E.3.)

We hope these additions answer the reviewer's questions, and we would be glad to provide further detail during the discussion period.

Best regards,

The Authors

Thank you for the controlled experiments and detailed clarifications. The matched-budget comparison addresses the original confound and supports an interruption–continuation trade-off rather than an overall macro advantage from mixing. The B-side example clarifies how stopping and responding are supervised, and the blinded listening study provides independent evidence for the adaptation gains.

Two reporting details would benefit from clarification:

  1. The original Moshi baseline changes from 26.43% macro in the submission to 24.42% in the controlled experiment. What accounts for this difference, and how should the two sets of results be compared?
  2. For M-MOS, the reported means imply a difference of 0.32 (4.41 − 4.09), whereas the paired difference is 0.35. Please verify the values or clarify whether different aggregation was used.
AuthorsOpenReview ↗

Response to Reviewer crWP

Dear Reviewer crWP,

Thank you for the careful reading. We are glad the additions were helpful.

  1. Original Moshi baseline: 26.43% versus 24.42%.

The original checkpoint scored 26.43 in the submission run and 24.42 when we reran it for the controlled experiment in Table 5. Both runs used the same 498 inputs, decoding settings, ASR, and judge snapshot. They differed in the generated outputs: generations are sampled by specific temperatures and top-k settings, so scores can vary between runs. The task scores changed from 59.50 to 58.50 for interruption, 11.22 to 8.16 for backchannel, and 14.00 to 10.00 for background; talking-to-other was unchanged. The comparisons in Table 5 use 24.42 as the reference, evaluated in the same experimental run as the other rows. Table 6 uses the submission result, 26.43. We therefore compare results against the reference in their own table, rather than comparing absolute scores across the two tables.

  1. M-MOS means and paired difference.

For M-MOS 0.32 and 0.35, Both values are correct; they aggregate different sets. The system columns average all valid ratings (original 116, adapted 120 of 120), while the paired difference — the prespecified estimate — uses only the 116 rater-and-condition-matched pairs, averaged within condition and then over the 40 conditions. The four adapted-system ratings without an original counterpart (three conditions one rater marked N/A for silence in the original, and one dialogue another rater did not complete) score 3, 4, 3, and 5 — below the adapted average — so including them lowers the adapted column mean while the paired difference remains +0.35.

We apologize for the confusion these two reporting details caused. In the final version we will make both explicit: the reference row of Table 5 will be clearly labeled as re-evaluated within the controlled experiment's run, with the caption cross-noting the submission value, and the Table 3 caption will state that column means use all valid ratings while paired differences use matched pairs only. The current revision already carries these notes in Table 5's caption, Table 3's caption, and Appendices E.1 and E.3.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global response: the matched-budget experiment

We thank both reviewers for converging on the same central issue: in Table 13 of the original submission (now Table 16), synthesis route, data volume, and optimization budget vary together, so the comparison supports recipe-level utility but does not isolate the effect of mixing. We agree that the original comparison confounds data composition and training budget, so we ran the controlled experiment as follows.

Experimental design. Rule-only vs. mixed training at identical data and optimization budgets (113.07 h, 2,000 updates, shared initialization and validation), crossed with original vs. onset-corrected timing and three training seeds — twelve runs, each evaluated on all 498 Full-Duplex-Bench v1.5 cases under one frozen pipeline, with the same judge as the paper and all unknown labels kept in the denominator. Full details are provided in Section 6 and Appendix E.1 of the revision.

Results (three-seed means; macro with 95% case-bootstrap CI):

Arm (113.07 h)InterruptionBackchannelTalking-to-otherBackgroundMacro
Rule-only, original timing38.6729.2553.6749.0042.65 [39.1, 46.1]
Mixed, original timing59.0022.4530.3324.0033.95 [30.8, 37.1]
Rule-only, corrected timing30.6724.1552.3343.0037.54 [34.0, 41.0]
Mixed, corrected timing55.1719.7333.3327.0033.81 [30.5, 37.1]
Original Moshi (no fine-tuning, re-eval.)58.508.1621.0010.0024.42

What changes in the paper. At matched budget, mixing does not improve the equal-task macro (Mixed − Rule: −8.70 pp [−12.5, −4.8]); the original +16.6-point result is now presented as recipe-level utility only. Mixed training improves interruption response (+20.33 pp [+14.3, +26.2]), with a larger gain after timing correction, but produces lower continuation scores on the continue-speaking tasks. The revision states this trade-off explicitly (Section 6, Table 5, p. 9; Appendix E.1, Table 20). We are grateful for this line of questioning: it replaced our original claim with a more precise one.

AuthorsOpenReview ↗

Summary of changes during the discussion period

Dear AC and reviewers,

Since the discussion is closing soon, here is a short summary of what changed, mainly so that everything is in one place.

The main issue in both reviews was the same one: in the original Table 13, the native, rule-only and mixed conditions differed in data volume and training budget as well as in synthesis route, so the +16.6-point mixed-data result could not be attributed to mixing itself. We agreed and ran a controlled version: rule-only vs. mixed at the same 113.07 h and 2,000 updates, with and without the timing correction from Appendix B.3, three seeds each (12 runs). At matched budget, mixing gives a large gain on interruption response (+24.50 pp after correction) but a lower equal-task macro (−3.73 pp). The paper now reports this trade-off directly and describes the original result as recipe-level (Section 6, Table 5, Table 20, Appendix E.1). Reviewer crWP wrote that the matched-budget comparison addresses the original confound and raised their soundness score from 2 to 3.

Reviewer u4Wh asked for three further things, and all are in the revision: a comparison table with Lychee-FD[1], DuplexGen[2], ConversationalVoice[3], DuplexChat[4] and SmoothConv&DuplexConv[5] (Table 1, with source notes in Appendix E.6); an evaluation of the released Lychee-FD model on our 498-case protocol, where it reaches 27.03 macro against 33.81–42.65 for the matched DuplexTTS arms (Table 5, Appendix E.2); and event counts for every rule subset, which show that the matched arms have the same event mixture as the original corpus (Table 21). We also posted these tables inline in the reviewer's thread on Nov 17.

For Reviewer crWP we added a worked example of how B's reply is supervised after an interruption (Figure 8, Appendix E.4), a blind paired listening study comparing the original and adapted generator (Table 3; the three dialogue-level scores improved by 0.35 to 0.69, audio quality nearly unchanged), and two caption clarifications about the re-evaluated baseline and the paired MOS aggregation.

We have not yet heard back from Reviewer u4Wh. Since the central weakness in their review is the same concern that crWP considered addressed, and their other requests have been met, we hope the additions can be taken into account. We would be glad to answer any questions that comes up before the deadline.

We sincerely thank both reviewers for taking the time to carefully read our paper and provide thoughtful feedback.

Best regards,

The Authors

[1] Liu Z, Zhang X, Li Y, et al. Hierarchical Acoustic-Semantic Modeling: Modality Separation and Semantic Coherence for Full-Duplex SLMs[C]//Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026: 9264-9280.

[2] Kim T, Kim K, Woo T, et al. DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues[C]//Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing. Budapest, Hungary: Association for Computational Linguistics, 2026: 59479–59505. DOI: 10.18653/2026.emnlp-main.2618.

[3] He R Y, Cao B, Xu C, et al. ConversationalVoice: Full-Duplex Speech Data from Real Conversations through Source-Faithful Reconstruction and Conversation-Grounded Expansion[J]. arXiv preprint arXiv:2609.08147, 2026.

[4] Nakata W, Saito Y, Saruwatari H. DuplexChat: Constructing Speaker-Separated Full-Duplex Dialogue Speech at Scale for Spoken Dialogue Language Modeling[J]. arXiv preprint arXiv:2607.04941, 2026.

[5] Wang C, Shao M, He C, et al. SmoothConv and DuplexConv: Complementary Mandarin Multi-Party Conversational Speech Corpora for Speech Interaction[J]. arXiv preprint arXiv:2610.11150, 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.