-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
1337 lines (1282 loc) · 54.7 KB
/
Copy pathindex.html
File metadata and controls
1337 lines (1282 loc) · 54.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="description"
content="DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues." />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<meta property="og:title" content="DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues" />
<meta property="og:description"
content="A human-calibrated framework that generates synthetic human-AI dialogues with scenario-adaptive turn-taking." />
<title>DuplexGen</title>
<link rel="icon" href="data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'%3E%3Ctext y='.9em' font-size='90'%3E📻%3C/text%3E%3C/svg%3E" />
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet" />
<link rel="stylesheet" href="static/css/bulma.min.css" />
<link rel="stylesheet" href="static/css/fontawesome.all.min.css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css" />
<link rel="stylesheet" href="static/css/index.css" />
<script defer src="static/js/fontawesome.all.min.js"></script>
<script src="static/js/wavesurfer.js"></script>
<script src="static/js/sample_transcripts.js"></script>
<style>
:root {
--theme: #2f855a;
--theme-dark: #246b48;
--theme-muted: #78a98a;
--theme-soft: #edf8f1;
--theme-soft-strong: #e2f2e8;
--theme-border: #cfe5d7;
--theme-focus: rgba(47, 133, 90, 0.32);
--section-tint: #fafafa;
}
html { scroll-behavior: smooth; }
a:not(.button) { color: var(--theme); }
a:not(.button):hover { color: var(--theme-dark); }
.publication-authors a { color: var(--theme) !important; }
.publication-links .button.is-dark {
border-color: transparent;
background-color: var(--theme-dark);
}
.publication-links .button.is-dark:hover {
background-color: var(--theme);
}
.publication-links .button.is-dark:focus-visible {
outline: 3px solid var(--theme-focus);
outline-offset: 3px;
}
.section-tint,
.footer { background-color: var(--section-tint); }
.author-block { margin-right: 0.75rem; }
.author-list,
.affiliation-list {
display: flex;
flex-wrap: wrap;
justify-content: center;
}
.author-list { gap: 0.15rem 0.55rem; }
.affiliation-list { gap: 0.2rem 1rem; }
.affiliation-break {
flex-basis: 100%;
height: 0;
}
.author-list .author-block,
.affiliation-list .author-block { margin-right: 0; }
@media screen and (min-width: 769px) {
.author-list {
display: grid;
grid-template-columns: repeat(3, max-content);
gap: 0.15rem 0.8rem;
justify-content: center;
}
}
.section-nav {
position: sticky;
top: 0;
z-index: 40;
display: flex;
flex-wrap: nowrap;
justify-content: center;
gap: 0.45rem;
margin: 0;
padding: 0.55rem 1rem;
overflow-x: auto;
border-bottom: 1px solid rgba(207, 229, 215, 0.92);
background: rgba(255, 255, 255, 0.94);
backdrop-filter: blur(10px);
scrollbar-width: none;
}
.section-nav::-webkit-scrollbar { display: none; }
.section-nav a {
flex: none;
padding: 0.35rem 0.8rem;
border: 1px solid var(--theme-border);
border-radius: 999px;
background: #fff;
color: var(--theme);
font-size: 0.82rem;
font-weight: 600;
}
.section-nav a:hover {
border-color: var(--theme);
background: var(--theme-soft);
}
.section-nav a.is-active {
border-color: var(--theme);
background: var(--theme);
color: #fff;
}
.section-nav a:focus-visible {
outline: 3px solid var(--theme-focus);
outline-offset: 2px;
}
section[id] { scroll-margin-top: 4.5rem; }
.scenario-card { height: 100%; }
.scenario-tag { font-weight: 600; }
.waveform { width: 100%; }
.sample-title { margin-bottom: 0.25rem; }
.stage-number {
display: inline-flex; align-items: center; justify-content: center;
width: 2rem; height: 2rem; border-radius: 999px;
background: var(--theme); color: #fff; font-weight: 700; margin-right: 0.6rem;
}
.pipeline-stage { margin-bottom: 1rem; }
.overview-figure { margin: 1.75rem 0 0; text-align: center; }
.overview-figure img {
width: 100%; height: auto; background: #fff;
border: 1px solid #e6e6e6; border-radius: 6px;
}
.overview-figure figcaption {
font-size: 0.78rem; line-height: 1.5; color: #666; margin-top: 0.6rem;
}
.motivation-callout {
margin-top: 1.25rem;
padding: 1.35rem 1.5rem;
background: var(--theme-soft);
border: 1px solid var(--theme-border);
border-left: 4px solid var(--theme);
border-radius: 8px;
box-shadow: 0 3px 12px rgba(47, 133, 90, 0.08);
text-align: left;
}
.motivation-callout p {
margin: 0;
font-size: 1.05rem;
line-height: 1.75;
}
.motivation-callout strong {
color: var(--theme-dark);
}
.table-scroll-hint {
display: none;
margin: 1rem 0 0.35rem;
color: #666;
font-size: 0.78rem;
text-align: center;
}
.performance-table-wrap {
margin-top: 1.5rem;
overflow-x: auto;
overscroll-behavior-inline: contain;
padding-bottom: 0.25rem;
-webkit-overflow-scrolling: touch;
}
.performance-table-wrap:focus-visible {
outline: 3px solid var(--theme-focus);
outline-offset: 3px;
}
.performance-table {
min-width: 840px;
table-layout: fixed;
font-size: 0.9rem;
}
.performance-table caption {
caption-side: bottom;
padding: 0.85rem 0.5rem 0;
color: #666;
font-size: 0.8rem;
line-height: 1.5;
text-align: left;
}
.performance-table thead th,
.performance-table tbody th,
.performance-table tbody td {
padding: 0.55rem 0.65rem;
text-align: center !important;
vertical-align: middle;
}
.performance-table thead th {
color: #363636;
}
.performance-table .model-heading {
display: inline-flex;
min-height: 2.5em;
align-items: center;
justify-content: center;
line-height: 1.25;
white-space: nowrap;
}
.performance-table .model-heading.is-stacked {
flex-direction: column;
}
.performance-table .metric-group {
border-bottom: 2px solid var(--theme);
}
.performance-table .naturalness-divider {
border-left: 2px solid #cfd6df;
}
.performance-table .scenario {
width: 5rem;
font-weight: 700;
text-align: center !important;
}
.performance-table .mean,
.performance-table .sem {
display: block;
width: 100%;
white-space: nowrap;
text-align: center;
}
.performance-table .sem {
margin-top: 0.08rem;
color: #777;
font-size: 0.72rem;
font-weight: 400;
}
.performance-table tr.section-break > * {
border-top: 2px solid #cfd6df;
}
.performance-table tr.average > * {
border-top: 2px solid #7a7a7a;
background: #fafafa;
}
.performance-table .duplexgen-column {
background: var(--theme-soft);
}
.performance-table tbody tr:hover > * {
background: #f5faf7;
}
.performance-table tbody tr:hover > .duplexgen-column {
background: var(--theme-soft-strong);
}
.performance-table thead .scenario,
.performance-table tbody .scenario {
position: sticky;
left: 0;
z-index: 5;
background: #fff;
box-shadow: 2px 0 0 #dfe9e3;
}
.performance-table thead .scenario { z-index: 6; }
.performance-table tbody .scenario:hover,
.performance-table tbody .scenario:focus-within { z-index: 30; }
.performance-table tbody tr:hover > .scenario { background: #f5faf7; }
.performance-table tbody tr.average > .scenario { background: #fafafa; }
.performance-table .scenario abbr {
position: relative;
display: inline-block;
border-bottom: 1px dotted #777;
cursor: default;
text-decoration: none;
text-underline-offset: 3px;
}
.performance-table .scenario abbr::after {
content: attr(data-tooltip);
position: absolute;
z-index: 20;
top: calc(100% + 0.45rem);
left: 0;
visibility: hidden;
padding: 0.35rem 0.55rem;
border-radius: 5px;
background: #363636;
box-shadow: 0 3px 10px rgba(0, 0, 0, 0.18);
color: #fff;
font-size: 0.72rem;
font-weight: 600;
line-height: 1.25;
opacity: 0;
pointer-events: none;
transform: translateY(-0.2rem);
transition: opacity 0.15s ease, transform 0.15s ease, visibility 0.15s;
white-space: nowrap;
}
.performance-table .scenario abbr:hover::after,
.performance-table .scenario abbr:focus-visible::after {
visibility: visible;
opacity: 1;
transform: translateY(0);
}
.performance-table .scenario abbr:focus-visible {
border-bottom-color: transparent;
outline: 3px solid var(--theme-focus);
outline-offset: 2px;
}
.cmp-card {
background: #fff; border: 1px solid #e6e6e6; border-radius: 8px;
padding: 0.9rem; height: 100%; display: flex; flex-direction: column;
}
.wave-play {
display: block; width: 100%; margin-top: 0.4rem; border: none;
background: var(--theme); color: #fff; border-radius: 6px;
padding: 0.3rem 0.8rem; cursor: pointer; font-size: 0.85rem;
}
.wave-play:hover { background: var(--theme-dark); }
.wave-play:disabled {
background: var(--theme-muted);
cursor: wait;
opacity: 0.85;
}
.wave-play.is-error:disabled {
background: #8a8a8a;
cursor: not-allowed;
}
.transcript-toggle {
display: block;
width: 100%;
margin-top: 0.5rem;
padding: 0.3rem 0.8rem;
border: 1px solid var(--theme-border);
border-radius: 6px;
background: #fff;
color: var(--theme);
cursor: pointer;
font-size: 0.82rem;
font-weight: 600;
}
.transcript-toggle:hover { background: var(--theme-soft); }
.transcript-toggle:focus-visible {
outline: 3px solid var(--theme-focus);
outline-offset: 2px;
}
/* Selectable text transcript (bubbles) */
.cmp-transcript {
margin-top: 0.7rem; border-top: 1px solid #eee; padding-top: 0.6rem;
}
.cmp-transcript[hidden] { display: none; }
.cmp-transcript > .tr-caption { font-size: 0.7rem; color: #999; margin-bottom: 0.5rem; }
.cmp-tr { max-height: 460px; overflow-y: auto; padding-right: 0.3rem; }
.cmp-turn { display: flex; margin-bottom: 0.5rem; }
.cmp-turn.assistant { justify-content: flex-end; }
.cmp-bubble {
max-width: 88%; border-radius: 12px; padding: 0.45rem 0.7rem;
border: 1px solid #ccc; font-size: 0.82rem; line-height: 1.35;
}
.cmp-turn.human .cmp-bubble { background: #fff; border-color: #d2d2d2; }
.cmp-turn.assistant .cmp-bubble { background: #ebebeb; border-color: #c2c2c2; }
.cmp-bubble .who {
display: block; font-size: 0.66rem; font-weight: 700; color: #6e6e6e;
margin-bottom: 0.15rem; letter-spacing: 0.01em;
}
.cmp-bubble .txt { color: #1a1a1a; }
.cmp-tr-empty { font-size: 0.8rem; color: #b0392f; font-style: italic; padding: 0.4rem 0; }
/* Channel color-coding: human = red, AI = blue,
matching the dual-channel waveform colors. */
.lbl-ai { color: #1976D2; font-weight: 700; }
.lbl-human { color: #D32F2F; font-weight: 700; }
.cmp-bubble .who.lbl-ai { color: #1976D2; }
.cmp-bubble .who.lbl-human { color: #D32F2F; }
.bibtex-heading {
display: flex;
align-items: center;
justify-content: space-between;
gap: 1rem;
margin-bottom: 1rem;
}
.bibtex-heading .title { margin-bottom: 0; }
.copy-bibtex {
flex: none;
padding: 0.45rem 0.8rem;
border: 1px solid var(--theme-border);
border-radius: 6px;
background: #fff;
color: var(--theme);
cursor: pointer;
font-size: 0.82rem;
font-weight: 700;
}
.copy-bibtex:hover { background: var(--theme-soft); }
.copy-bibtex.is-copied {
border-color: #3a9d67;
background: #effaf4;
color: #26734a;
}
.copy-bibtex:focus-visible {
outline: 3px solid var(--theme-focus);
outline-offset: 2px;
}
.visually-hidden {
position: absolute !important;
width: 1px !important;
height: 1px !important;
padding: 0 !important;
margin: -1px !important;
overflow: hidden !important;
clip: rect(0, 0, 0, 0) !important;
white-space: nowrap !important;
border: 0 !important;
}
.back-to-top {
position: fixed;
right: max(1rem, env(safe-area-inset-right));
bottom: max(1rem, env(safe-area-inset-bottom));
z-index: 50;
display: inline-flex;
align-items: center;
gap: 0.35rem;
padding: 0.55rem 0.75rem;
border: none;
border-radius: 999px;
background: var(--theme);
box-shadow: 0 5px 16px rgba(31, 92, 59, 0.25);
color: #fff;
cursor: pointer;
font-size: 0.82rem;
font-weight: 700;
opacity: 0;
pointer-events: none;
transform: translateY(0.75rem);
transition: opacity 0.2s ease, transform 0.2s ease, visibility 0.2s;
visibility: hidden;
}
.back-to-top.is-visible {
opacity: 1;
pointer-events: auto;
transform: translateY(0);
visibility: visible;
}
.back-to-top.is-visible:hover {
background: var(--theme-dark);
transform: translateY(-2px);
}
.back-to-top:focus-visible {
outline: 3px solid var(--theme-focus);
outline-offset: 3px;
}
@media screen and (max-width: 768px) {
.section-nav { justify-content: flex-start; }
.author-list { gap: 0.1rem 0.4rem; }
.affiliation-list {
gap: 0.15rem 0.65rem;
font-size: 0.85rem !important;
line-height: 1.45;
}
.affiliation-break { display: none; }
.motivation-callout {
padding: 1.1rem 1.15rem;
}
.table-scroll-hint {
display: block;
}
.performance-table-wrap {
margin-top: 0;
}
.has-text-justified {
text-align: left !important;
}
.bibtex-heading {
align-items: flex-start;
}
}
@media (prefers-reduced-motion: reduce) {
html { scroll-behavior: auto; }
.back-to-top { transition: none; }
.performance-table .scenario abbr::after { transition: none; }
}
</style>
</head>
<body>
<!-- Hero -->
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">DuplexGen</h1>
<h2 class="subtitle is-3">Adaptive Synthesis of Human–AI Turn-Taking Dialogues</h2>
<div class="is-size-5 publication-authors author-list" style="margin-top:1rem;">
<span class="author-block"><a href="https://youngerous.github.io/" target="_blank" rel="noopener">Takyoung Kim</a><sup>1</sup><sup>*</sup>,</span>
<span class="author-block"><a href="https://kwkim.me/" target="_blank" rel="noopener">Kang-wook Kim</a><sup>2,4</sup><sup>*</sup>,</span>
<span class="author-block"><a href="https://tonywoo.me/" target="_blank" rel="noopener">Sang Hoon Woo</a><sup>2,5</sup>,</span>
<span class="author-block"><a href="https://www.cs.columbia.edu/~julia/" target="_blank" rel="noopener">Julia Hirschberg</a><sup>3</sup>,</span>
<span class="author-block"><a href="https://vision.snu.ac.kr/gunhee/" target="_blank" rel="noopener">Gunhee Kim</a><sup>2</sup>,</span>
<span class="author-block"><a href="https://siebelschool.illinois.edu/about/people/faculty/dilek" target="_blank" rel="noopener">Dilek Hakkani-Tür</a><sup>1</sup></span>
</div>
<div class="is-size-6 publication-authors affiliation-list" style="margin-top:0.5rem;">
<span class="author-block"><sup>1</sup>University of Illinois Urbana-Champaign</span>
<span class="author-block"><sup>2</sup>Seoul National University</span>
<span class="author-block"><sup>3</sup>Columbia University</span>
<span class="affiliation-break" aria-hidden="true"></span>
<span class="author-block"><sup>4</sup>University of California, Berkeley</span>
<span class="author-block"><sup>5</sup>Georgia Institute of Technology</span>
</div>
<div class="is-size-6 publication-authors" style="margin-top:0.35rem;">
<span class="author-block"><sup>*</sup>Equal contribution</span>
</div>
<div class="is-size-5 publication-authors" style="margin-top:0.35rem;">
<span class="author-block"><strong>EMNLP 2026</strong></span>
</div>
<!-- Resource buttons -->
<div class="column has-text-centered" style="margin-top:1.5rem;">
<div class="publication-links">
<span class="link-block">
<a href="https://arxiv.org/abs/2607.26178" target="_blank" rel="noopener"
class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fas fa-file-pdf"></i></span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/duplexgen/duplexgen-code"
target="_blank" rel="noopener"
class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fab fa-github"></i></span>
<span>Code</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/duplexgen/personaplex-finetune"
target="_blank" rel="noopener"
class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fas fa-microchip"></i></span>
<span>Finetune</span>
</a>
</span>
<span class="link-block">
<a href="https://huggingface.co/datasets/DuplexGen/duplexgen-corpus"
target="_blank" rel="noopener"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">🤗</span>
<span>Corpus</span>
</a>
</span>
<span class="link-block">
<a href="https://huggingface.co/datasets/DuplexGen/duplexgen-spoken"
target="_blank" rel="noopener"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">🤗</span>
<span>Spoken</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<nav class="section-nav" aria-label="Page sections">
<a href="#framework">Framework</a>
<a href="#scenarios">Scenarios</a>
<a href="#samples">Samples</a>
<a href="#evaluation">Evaluation</a>
</nav>
<!-- Abstract -->
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
Turn-taking is a central component of full-duplex interaction. Which turn-taking
behaviors are appropriate varies with the scenario, yet current models apply a single
norm regardless of context. This limitation originates in their training data:
human–human speech corpora capture natural timing phenomena but provide little
role grounding or scenario-specific norms, while heuristic or prompted synthesis
methods inject turn-taking behaviors without basing them on human preferences.
We introduce <strong>DuplexGen</strong>, a framework for generating dialogues with
scenario-adaptive turn-taking by calibrating LLM predictions against a small set of
slot-level human preference annotations.
In six cooperative and competitive tasks, human turn-taking preferences differ
systematically, and DuplexGen aligns substantially more closely with those
preferences than uncalibrated prompting or training solely on generic
human–human data; a full-duplex model trained on DuplexGen-generated data
exhibits distinctive, human-preferred turn-taking behaviors. These results show that
human calibration, not corpus scale or prompt design alone, is what allows
turn-taking synthesis to be scenario-specific.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Motivation -->
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Motivation: When Should an AI Speak?</h2>
<div class="content motivation-callout">
<p>
Imagine an <strong>AI tutor</strong> that interrupts whenever a student hesitates,
and an <strong>AI negotiator</strong> that patiently waits while the other side takes
control. Both systems may generate perfectly fluent responses, yet neither understands
<em>when</em> it should speak. Effective spoken AI must decide not only what to say,
but also when to listen, backchannel, or take the floor; the right decision depends
on its role.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Pipeline -->
<section class="section section-tint" id="framework">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">DuplexGen Framework</h2>
<p class="subtitle is-6 has-text-centered">
Four stages turn clean text dialogues into human-calibrated, scenario-adaptive
turn-taking dialogues.
</p>
<!-- Figure 1 of the paper. Regenerate from DuplexGen-EMNLP2026/assets/overview.pdf:
fitz.open(pdf)[0].get_pixmap(matrix=fitz.Matrix(2.8, 2.8), alpha=False), then
crop to the ink bounding box with 6 px of padding. -->
<figure class="overview-figure">
<!-- Links to the full-size render so the in-figure text stays readable on a phone. -->
<a href="static/images/overview.png" target="_blank" rel="noopener">
<img src="static/images/overview.png"
alt="Four-panel overview of the DuplexGen framework: Convert (spoken-style dialogue
conversion), Identify (turn-taking slot identification), Calibrate (slot
annotation and calibration), Synthesize (applying the calibrated predictor to
target scenarios)." />
</a>
<figcaption>
A text dialogue goes in; a scenario-adapted turn-taking dialogue comes out. The four
stages are described below. Click to enlarge.
</figcaption>
</figure>
<div class="content" style="margin-top:1.5rem;">
<div class="pipeline-stage">
<span class="stage-number">1</span><strong>Spoken-Style Dialogue Conversion.</strong>
An LLM rewrites text-based dialogues into spoken-style transcripts, removing text-only
artifacts and adding natural disfluencies and fillers.
</div>
<div class="pipeline-stage">
<span class="stage-number">2</span><strong>Turn-Taking Slot Identification.</strong>
Heuristics and an LLM mark candidate intra-utterance action slots, the points
where the AI could take the floor, backchannel, or keep listening.
</div>
<div class="pipeline-stage">
<span class="stage-number">3</span><strong>Scenario-Specific Calibration.</strong>
A small set of slot-level human preference annotations calibrates the LLM's per-slot
action distribution via KL-divergence soft-label fine-tuning, so predictions reflect
human preference rather than raw model confidence.
</div>
<div class="pipeline-stage">
<span class="stage-number">4</span><strong>Scenario-Adaptive Dialogue Synthesis.</strong>
The predictor calibrated for the target scenario is applied slot-by-slot to insert the
AI's turn-taking behavior: <strong>floor-taking</strong>,
<strong>backchanneling</strong>, or <strong>silence</strong>, yielding a full turn-taking
dialogue.
</div>
<p class="is-size-7" style="margin-top:0.75rem;color:#666;">
For listening tests, synthesized dialogues are rendered to full-duplex speech with
<a href="https://github.com/resemble-ai/chatterbox" target="_blank"
rel="noopener">Chatterbox</a> TTS.
</p>
</div>
</div>
</section>
<!-- Scenario taxonomy -->
<section class="section" id="scenarios">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Six Conversational Scenarios</h2>
<p class="subtitle is-6 has-text-centered">
Cooperative and competitive human–AI tasks, each grounded in an established dialogue
dataset. Turn-taking norms differ significantly across them.
</p>
<h3 class="title is-5 has-text-info" style="margin-top:1rem;">Cooperative</h3>
<div class="columns">
<div class="column">
<div class="box scenario-card">
<p class="scenario-tag has-text-info">TEA: Socratic Teaching</p>
<p>Guided knowledge discovery, where the tutor leads a student through a problem
rather than giving the answer.</p>
<p class="is-size-7 has-text-grey">Source: SocraticLM</p>
</div>
</div>
<div class="column">
<div class="box scenario-card">
<p class="scenario-tag has-text-info">PLN: Mixed-Initiative Planning</p>
<p>Collaborative goal completion, gathering information and comparing options toward
a plan.</p>
<p class="is-size-7 has-text-grey">Source: MultiWOZ</p>
</div>
</div>
<div class="column">
<div class="box scenario-card">
<p class="scenario-tag has-text-info">INT: Interview</p>
<p>Structured information elicitation through open-ended questions and follow-up
probing.</p>
<p class="is-size-7 has-text-grey">Source: Anthropic Interviewer</p>
</div>
</div>
</div>
<h3 class="title is-5 has-text-danger" style="margin-top:1rem;">Competitive</h3>
<div class="columns">
<div class="column">
<div class="box scenario-card">
<p class="scenario-tag has-text-danger">NEG: Negotiation</p>
<p>Resource allocation under conflict: bargaining over price and terms toward
a deal.</p>
<p class="is-size-7 has-text-grey">Source: CraigslistBargain</p>
</div>
</div>
<div class="column">
<div class="box scenario-card">
<p class="scenario-tag has-text-danger">PER: Persuasion</p>
<p>Attitude-change attempts, steering the other party toward a position or action.</p>
<p class="is-size-7 has-text-grey">Source: DailyPersuasion</p>
</div>
</div>
<div class="column">
<div class="box scenario-card">
<p class="scenario-tag has-text-danger">SOC: Social Chat</p>
<p>Peer-like conversation with everyday, emotionally charged give-and-take.</p>
<p class="is-size-7 has-text-grey">Source: SODA</p>
</div>
</div>
</div>
</div>
</section>
<!-- Audio samples -->
<section class="section section-tint" id="samples">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">DuplexGen Corpus Samples</h2>
<p class="subtitle is-6 has-text-centered">
One DuplexGen-synthesized full-duplex dialogue per scenario, rendered with Chatterbox TTS.
The full corpus is on the
<a href="https://huggingface.co/datasets/DuplexGen/duplexgen-spoken"
target="_blank" rel="noopener">Hub</a>.
</p>
<div class="columns is-multiline" style="margin-top:1rem;">
<div class="column is-half">
<div class="cmp-card">
<p class="sample-title"><strong>TEA</strong>: Socratic Teaching</p>
<div id="wave_TEA" class="waveform"></div>
<button class="transcript-toggle" type="button" aria-expanded="false"
aria-controls="transcript_TEA">Show transcript</button>
<div id="transcript_TEA" class="cmp-transcript" hidden><div id="tr_TEA"></div></div>
</div>
</div>
<div class="column is-half">
<div class="cmp-card">
<p class="sample-title"><strong>PLN</strong>: Mixed-Initiative Planning</p>
<div id="wave_PLN" class="waveform"></div>
<button class="transcript-toggle" type="button" aria-expanded="false"
aria-controls="transcript_PLN">Show transcript</button>
<div id="transcript_PLN" class="cmp-transcript" hidden><div id="tr_PLN"></div></div>
</div>
</div>
<div class="column is-half">
<div class="cmp-card">
<p class="sample-title"><strong>INT</strong>: Interview</p>
<div id="wave_INT" class="waveform"></div>
<button class="transcript-toggle" type="button" aria-expanded="false"
aria-controls="transcript_INT">Show transcript</button>
<div id="transcript_INT" class="cmp-transcript" hidden><div id="tr_INT"></div></div>
</div>
</div>
<div class="column is-half">
<div class="cmp-card">
<p class="sample-title"><strong>NEG</strong>: Negotiation</p>
<div id="wave_NEG" class="waveform"></div>
<button class="transcript-toggle" type="button" aria-expanded="false"
aria-controls="transcript_NEG">Show transcript</button>
<div id="transcript_NEG" class="cmp-transcript" hidden><div id="tr_NEG"></div></div>
</div>
</div>
<div class="column is-half">
<div class="cmp-card">
<p class="sample-title"><strong>PER</strong>: Persuasion</p>
<div id="wave_PER" class="waveform"></div>
<button class="transcript-toggle" type="button" aria-expanded="false"
aria-controls="transcript_PER">Show transcript</button>
<div id="transcript_PER" class="cmp-transcript" hidden><div id="tr_PER"></div></div>
</div>
</div>
<div class="column is-half">
<div class="cmp-card">
<p class="sample-title"><strong>SOC</strong>: Social Chat</p>
<div id="wave_SOC" class="waveform"></div>
<button class="transcript-toggle" type="button" aria-expanded="false"
aria-controls="transcript_SOC">Show transcript</button>
<div id="transcript_SOC" class="cmp-transcript" hidden><div id="tr_SOC"></div></div>
</div>
</div>
</div>
<p class="has-text-centered is-size-7" style="margin-top:0.5rem;">
Each waveform is dual-channel: <strong>top:</strong> <span class="lbl-human">human</span>,
<strong>bottom:</strong> <span class="lbl-ai">AI assistant</span>. Click a waveform to play.
</p>
</div>
</section>
<!-- Model comparison -->
<section class="section" id="evaluation">
<div class="container is-max-widescreen">
<h2 class="title is-3 has-text-centered">Full-Duplex Model Evaluation</h2>
<p id="performance-scroll-hint" class="table-scroll-hint">
Scroll horizontally to view all columns.
</p>
<div class="performance-table-wrap"
role="region"
aria-label="Full-duplex model evaluation results"
aria-describedby="performance-scroll-hint"
tabindex="0">
<table class="table is-fullwidth performance-table">
<caption id="performance-caption">
Multi-turn dialogue evaluation on a 5-point Likert scale. Instruction Following is
rated by <strong>GPT-4.1</strong> from transcripts; Turn-Taking Naturalness is rated
by <strong>humans</strong> via a listening test, since context-appropriate turn-taking
is hard for LLMs to judge but natural for humans. Values are mean ± standard
error of the mean. The best result in each scenario is shown in <strong>bold</strong>.
</caption>
<thead>
<tr>
<th class="scenario" scope="col" rowspan="2">Scenario</th>
<th class="metric-group" scope="colgroup" colspan="3">Instruction Following</th>
<th class="metric-group naturalness-divider" scope="colgroup" colspan="3">Turn-Taking Naturalness</th>
</tr>
<tr>
<th scope="col"><span class="model-heading">Moshi</span></th>
<th scope="col"><span class="model-heading">PersonaPlex</span></th>
<th class="duplexgen-column" scope="col">
<span class="model-heading is-stacked">
<span>PersonaPlex +</span>
<span>DuplexGen</span>
</span>
</th>
<th class="naturalness-divider" scope="col">
<span class="model-heading">Moshi</span>
</th>
<th scope="col"><span class="model-heading">PersonaPlex</span></th>
<th class="duplexgen-column" scope="col">
<span class="model-heading is-stacked">
<span>PersonaPlex +</span>
<span>DuplexGen</span>
</span>
</th>
</tr>
</thead>
<tbody>
<tr>
<th class="scenario" scope="row">
<abbr data-tooltip="Socratic Teaching"
aria-label="TEA: Socratic Teaching" tabindex="0">TEA</abbr>
</th>
<td><span class="mean">1.44</span><span class="sem">± 0.12</span></td>
<td><span class="mean">2.96</span><span class="sem">± 0.26</span></td>
<td class="duplexgen-column"><strong class="mean">3.40</strong><span class="sem">± 0.29</span></td>
<td class="naturalness-divider"><span class="mean">3.09</span><span class="sem">± 0.15</span></td>
<td><span class="mean">3.37</span><span class="sem">± 0.12</span></td>
<td class="duplexgen-column"><strong class="mean">3.75</strong><span class="sem">± 0.11</span></td>
</tr>
<tr>
<th class="scenario" scope="row">
<abbr data-tooltip="Mixed-Initiative Planning"
aria-label="PLN: Mixed-Initiative Planning" tabindex="0">PLN</abbr>
</th>
<td><span class="mean">1.40</span><span class="sem">± 0.12</span></td>
<td><strong class="mean">2.04</strong><span class="sem">± 0.16</span></td>
<td class="duplexgen-column"><span class="mean">1.72</span><span class="sem">± 0.14</span></td>
<td class="naturalness-divider"><span class="mean">3.59</span><span class="sem">± 0.13</span></td>
<td><span class="mean">3.73</span><span class="sem">± 0.10</span></td>
<td class="duplexgen-column"><strong class="mean">3.76</strong><span class="sem">± 0.09</span></td>
</tr>
<tr>
<th class="scenario" scope="row">
<abbr data-tooltip="Interview" aria-label="INT: Interview" tabindex="0">INT</abbr>
</th>
<td><span class="mean">3.96</span><span class="sem">± 0.23</span></td>
<td><span class="mean">3.76</span><span class="sem">± 0.28</span></td>
<td class="duplexgen-column"><strong class="mean">4.09</strong><span class="sem">± 0.32</span></td>
<td class="naturalness-divider"><strong class="mean">3.68</strong><span class="sem">± 0.10</span></td>
<td><span class="mean">3.56</span><span class="sem">± 0.11</span></td>
<td class="duplexgen-column"><span class="mean">3.56</span><span class="sem">± 0.12</span></td>
</tr>
<tr class="section-break">
<th class="scenario" scope="row">
<abbr data-tooltip="Negotiation" aria-label="NEG: Negotiation" tabindex="0">NEG</abbr>
</th>
<td><span class="mean">2.12</span><span class="sem">± 0.22</span></td>
<td><span class="mean">4.00</span><span class="sem">± 0.25</span></td>
<td class="duplexgen-column"><strong class="mean">4.12</strong><span class="sem">± 0.20</span></td>
<td class="naturalness-divider"><span class="mean">3.46</span><span class="sem">± 0.12</span></td>
<td><span class="mean">3.60</span><span class="sem">± 0.10</span></td>
<td class="duplexgen-column"><strong class="mean">3.63</strong><span class="sem">± 0.12</span></td>
</tr>
<tr>
<th class="scenario" scope="row">
<abbr data-tooltip="Persuasion" aria-label="PER: Persuasion" tabindex="0">PER</abbr>
</th>
<td><span class="mean">3.12</span><span class="sem">± 0.31</span></td>
<td><span class="mean">3.96</span><span class="sem">± 0.30</span></td>
<td class="duplexgen-column"><strong class="mean">4.12</strong><span class="sem">± 0.21</span></td>
<td class="naturalness-divider"><span class="mean">3.60</span><span class="sem">± 0.14</span></td>
<td><span class="mean">3.49</span><span class="sem">± 0.12</span></td>
<td class="duplexgen-column"><strong class="mean">3.75</strong><span class="sem">± 0.11</span></td>
</tr>
<tr>
<th class="scenario" scope="row">
<abbr data-tooltip="Social Chat" aria-label="SOC: Social Chat" tabindex="0">SOC</abbr>
</th>
<td><span class="mean">3.64</span><span class="sem">± 0.29</span></td>
<td><span class="mean">3.72</span><span class="sem">± 0.30</span></td>
<td class="duplexgen-column"><strong class="mean">3.92</strong><span class="sem">± 0.26</span></td>
<td class="naturalness-divider"><span class="mean">3.46</span><span class="sem">± 0.13</span></td>
<td><span class="mean">3.57</span><span class="sem">± 0.11</span></td>
<td class="duplexgen-column"><strong class="mean">3.69</strong><span class="sem">± 0.10</span></td>
</tr>
<tr class="average">
<th class="scenario" scope="row">Avg.</th>
<td><span class="mean">2.61</span><span class="sem">± 0.12</span></td>
<td><span class="mean">3.41</span><span class="sem">± 0.12</span></td>
<td class="duplexgen-column"><strong class="mean">3.55</strong><span class="sem">± 0.12</span></td>
<td class="naturalness-divider"><span class="mean">3.48</span><span class="sem">± 0.05</span></td>
<td><span class="mean">3.56</span><span class="sem">± 0.05</span></td>
<td class="duplexgen-column"><strong class="mean">3.69</strong><span class="sem">± 0.05</span></td>
</tr>
</tbody>
</table>
</div>
</div>
</section>
<!-- Results -->
<section class="section section-tint">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Key Findings</h2>
<!-- (background alternates: tinted samples, white evaluation, tinted results) -->
<div class="content has-text-justified">
<p>
<strong>Turn-taking preferences differ by scenario.</strong> Human annotations show
significantly different action distributions across nearly all scenario pairs:
backchanneling is more frequent in cooperative tasks, floor-taking in competitive ones.
</p>
<p>
<strong>Calibration recovers human-like turn-taking.</strong> DuplexGen's calibrated
predictions align far more closely with human slot-level preferences (lowest KL
divergence) than uncalibrated prompting or predictors trained solely on generic
human–human speech, which over-predict silence and under-predict backchannels.
</p>
<p>
<strong>Synthesized data shapes full-duplex behavior.</strong> A full-duplex model fine-tuned on DuplexGen dialogues (<strong>PersonaPlex + DuplexGen</strong>) exhibits
distinctive,
human-preferred turn-taking that adapts per scenario, and is rated significantly more
natural than the Moshi and PersonaPlex baselines.
</p>
<p>
See the paper for the
full evaluation protocol, baselines, and per-scenario numbers.
</p>
</div>
</div>
</section>
<!-- BibTeX -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<div class="bibtex-heading">
<h2 class="title is-3">BibTeX</h2>
<button id="copy-bibtex" class="copy-bibtex" type="button"
aria-describedby="bibtex-copy-status">Copy BibTeX</button>
</div>
<p id="bibtex-copy-status" class="visually-hidden" aria-live="polite"></p>
<pre><code id="bibtex-code">@misc{kim2026duplexgenadaptivesynthesishumanai,
title={DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues},
author={Takyoung Kim and Kang-wook Kim and Sang Hoon Woo and Julia Hirschberg and Gunhee Kim and Dilek Hakkani-Tür},
year={2026},
eprint={2607.26178},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2607.26178},
}</code></pre>
</div>