-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
870 lines (767 loc) · 33.9 KB
/
index.html
File metadata and controls
870 lines (767 loc) · 33.9 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description" content="MTalk-Bench: Evaluating Speech-to-Speech Models in Multi-Turn Dialogues">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>MTalk-Bench: Evaluating Speech-to-Speech Models in Multi-Turn Dialogues via Arena-style and Rubrics Protocols</title>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/bulma/0.9.4/css/bulma.min.css">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/academicons/1.9.4/css/academicons.min.css">
<style>
body {
font-family: 'Noto Sans', sans-serif;
}
.publication-title {
font-family: 'Google Sans', sans-serif;
}
.publication-authors {
font-family: 'Google Sans', sans-serif;
}
.publication-authors a {
color: hsl(204, 86%, 53%) !important;
}
.publication-authors a:hover {
text-decoration: underline;
}
.author-block {
display: inline-block;
}
.eql-cntrb {
font-size: smaller;
}
/* Custom Carousel Styles */
.carousel-container {
position: relative;
overflow: hidden;
width: 100%;
height: 500px; /* 减少高度 */
}
.carousel-wrapper {
display: flex;
transition: transform 0.5s ease-in-out;
height: 100%;
}
.carousel-item {
flex: 0 0 100%; /* 精确占满一屏，等价于 width: 100% 且不收缩 */
box-sizing: border-box; /* 把 padding 计算进 100% 宽度，避免溢出 */
display: flex;
align-items: center;
justify-content: center;
padding: 1rem;
}
.carousel-item figure {
width: 100%;
max-width: 900px; /* 增加最大宽度 */
margin: 0 auto;
}
.carousel-item img {
width: 100%;
height: auto;
border-radius: 12px;
box-shadow: 0 8px 32px rgba(0,0,0,0.1);
}
/* Navigation buttons */
.carousel-nav {
position: absolute;
top: 50%;
transform: translateY(-50%);
background: rgba(0,0,0,0.7);
color: white;
border: none;
width: 50px;
height: 50px;
border-radius: 50%;
cursor: pointer;
display: flex;
align-items: center;
justify-content: center;
font-size: 18px;
z-index: 10;
transition: background-color 0.3s ease;
}
.carousel-nav:hover {
background: rgba(0,0,0,0.9);
}
.carousel-nav-left {
left: 20px;
}
.carousel-nav-right {
right: 20px;
}
/* Pagination dots */
.carousel-pagination {
display: flex;
justify-content: center;
gap: 10px;
margin-top: 1rem;
}
.carousel-dot {
width: 12px;
height: 12px;
border-radius: 50%;
background: #ccc;
cursor: pointer;
transition: background-color 0.3s ease;
}
.carousel-dot.active {
background: #3273dc;
}
/* Demo cards styling */
.demo-card {
background: white;
border-radius: 16px;
padding: 2rem;
box-shadow: 0 4px 20px rgba(0,0,0,0.1);
height: 100%;
}
.demo-audio {
width: 100%;
margin-bottom: 0.5rem;
}
.set-block {
border-left: 3px solid #3273dc;
padding-left: 1rem;
margin: 1rem 0;
}
.demo-card .subcap {
border-radius: 9999px;
padding: 0.25rem 0.6rem;
font-weight: 600;
border: none;
}
/* 分场景配色：各自唯一颜色 */
.demo-card[data-demo="Semantic"] .subcap {
background-color: #2b8a3e; /* 绿色 */
color: #fff;
}
.demo-card[data-demo="Paralinguistic"] .subcap {
background-color: #6f42c1; /* 紫色 */
color: #fff;
}
.demo-card[data-demo="Ambient"] .subcap {
background-color: #1e88e5; /* 蓝色 */
color: #fff;
}
.key-findings li {
margin-bottom: 1rem;
padding: 0.5rem 0;
}
/* Results carousel specific styling - 调整布局使caption在图片下方 */
.results-carousel-container {
height: 600px; /* 减少高度 */
}
.results-carousel-item .card {
height: 100%;
display: flex;
flex-direction: column;
}
.results-carousel-item .card-content {
flex: 1;
display: flex;
flex-direction: column;
padding: 0.5rem; /* 减少内边距 */
}
.results-carousel-item figure {
flex: 1;
display: flex;
flex-direction: column; /* 改为纵向布局 */
justify-content: center;
align-items: center;
margin: 0; /* 移除默认边距 */
}
.results-carousel-item figure img {
width: 100%;
height: auto;
max-height: 450px; /* 减少图片最大高度 */
object-fit: contain;
}
.figure-caption,
.image-caption {
margin-top: 0.75rem; /* 减少上边距 */
margin-bottom: 0; /* 移除下边距 */
font-size: 0.95rem;
line-height: 1.4;
color: rgba(0,0,0,0.7);
text-align: center;
width: 100%;
}
/* 减少section之间的间距 */
.section {
padding: 2rem 1.5rem; /* 减少section的padding */
}
.hero.is-small .hero-body {
padding: 1.5rem 1.5rem; /* 减少hero的padding */
}
/* Compact spacing for the two findings sections */
#results-analysis .content,
#meta-eval .content {
line-height: 1.5; /* 收紧段内行距 */
}
#results-analysis .key-findings,
#meta-eval .key-findings {
margin-top: 0.25rem; /* 收紧列表上下空白 */
margin-bottom: 0.75rem;
}
#results-analysis .key-findings li,
#meta-eval .key-findings li {
margin-bottom: 0.35rem; /* 收紧条目间距（原来是 1rem） */
padding: 0.15rem 0; /* 收紧条目内边距（原来是 0.5rem 0） */
line-height: 1.45; /* 子弹条目文本行距更紧一点 */
}
</style>
</head>
<body>
<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">MTalk-Bench: Evaluating Speech-to-Speech Models in Multi-Turn Dialogues via Arena-style and Rubrics Protocols</h1>
<div class="is-size-5 publication-authors" style="margin-top: 1em;">
Yuhao Du<sup>*</sup>
Qianwei Huang<sup>*</sup>
Guo Zhu
Zhanchen Dai
Shunian Chen<br>
Qiming Zhu
Yuhao Zhang
Li Zhou
Benyou Wang<sup>†</sup>
Haizhou Li<sup>†</sup>
</div>
<div class="is-size-6 publication-authors">
<span class="author-block">
School of Data Science, The Chinese University of Hong Kong, Shenzhen
</span><br>
<a href="https://github.com/FreedomIntelligence/MTalk-Bench" target="_blank">
https://github.com/FreedomIntelligence/MTalk-Bench
</a>
</div>
<div class="is-size-6 eql-cntrb" style="margin-top: 0.75em;">
<small>
<sup>*</sup> Equal contribution. <sup>†</sup> Corresponding author.
</small>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a href="#" target="_blank" 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://huggingface.co/datasets/FreedomIntelligence/MTalk-Bench" target="_blank" class="external-link button is-normal is-rounded is-dark">
<span>🤗 Dataset</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/FreedomIntelligence/MTalk-Bench" target="_blank" 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>
<!-- ArXiv abstract Link -->
<span class="link-block">
<a href="https://arxiv.org/abs/<ARXIV PAPER ID>" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Paper abstract -->
<section class="section hero is-light">
<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>
The rapid advancement of speech-to-speech (S2S) large language models (LLMs) has significantly improved real-time spoken interaction. However, current evaluation frameworks remain inadequate for assessing performance in complex, multi-turn dialogues.
To address this, we introduce <strong>MTalk-Bench</strong>, a multi-turn S2S benchmark covering three core dimensions:
<em>Semantic Information</em>, <em>Paralinguistic Information</em>, and <em>Ambient Sound</em>.
Each dimension includes nine realistic scenarios, along with targeted tasks to assess specific capabilities such as reasoning.
Our dual-method evaluation framework combines <em>Arena-style</em> evaluation (pairwise comparison) and <em>Rubrics-based</em>
evaluation (absolute scoring) for relative and absolute assessment. The benchmark includes both model and human outputs, evaluated by
human evaluators and LLMs.
Experimental results reveal two sets of findings. <strong>Overall performance of S2S LLMs:</strong>
(1) models excel at semantic information processing yet underperform on paralinguistic information and ambient sounds perception;
(2) Models typically regain coherence by increasing response length, sacrificing efficiency in multi-turn dialogues;
(3) modality-aware, task-specific designs outperform brute scaling.
<strong>Evaluation framework and reliability:</strong>
(1) Arena and Rubrics yield consistent, complementary rankings, but reliable distinctions emerge only when performance gaps are large;
(2) LLM-as-a-judge aligns with humans when gaps are clear or criteria explicit, but exhibits position and length biases and is reliable
on nonverbal evaluation only with text annotations.
These results highlight current limitations in S2S evaluation and the need for more robust, speech-aware assessment frameworks.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Framework & Evaluation (Condensed, English-only) -->
<section class="section" id="framework-eval">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Framework & Evaluation</h2>
<!-- Figures: concise academic captions -->
<section class="hero is-small" id="figures">
<div class="hero-body">
<div class="container">
<div class="carousel-container">
<div class="carousel-wrapper" id="fig-carousel-wrapper">
<!-- Slide 1 -->
<div class="carousel-item">
<figure class="image">
<img src="static/images/main_structure_comics_new.png" alt="Framework overview" loading="lazy">
<figcaption class="figure-caption">
<strong>Figure 1. MTalk-Bench overview.</strong> Three speech-aware dimensions
(Semantic, Paralinguistic, Ambient) organized into a hierarchical capability taxonomy,
evaluated via a <em>dual-method</em> design: Arena-style and rubric-based evaluation,
executed by both human raters and LLM-as-a-judge.
</figcaption>
</figure>
</div>
<!-- Slide 2 -->
<div class="carousel-item">
<figure class="image">
<img src="static/images/Data_construction_comics_new.png" alt="Data pipeline" loading="lazy">
<figcaption class="figure-caption">
<strong>Figure 2. Data construction pipeline.</strong> Scenario-capability mapping → multi-turn script authoring →
human speech recording → Seed-VC timbre conversion (for child/elder voices) → ambient-sound integration → quality checking → inputs to Arena & Rubrics evaluations.
</figcaption>
</figure>
</div>
</div>
<!-- Navigation -->
<button class="carousel-nav carousel-nav-left" id="fig-prev" aria-label="Previous figure">
<i class="fas fa-chevron-left" aria-hidden="true"></i>
</button>
<button class="carousel-nav carousel-nav-right" id="fig-next" aria-label="Next figure">
<i class="fas fa-chevron-right" aria-hidden="true"></i>
</button>
</div>
<!-- Pagination -->
<div class="carousel-pagination" id="fig-pagination" aria-label="Figures pagination">
<span class="carousel-dot active" data-slide="0"></span>
<span class="carousel-dot" data-slide="1"></span>
</div>
</div>
</div>
</section>
<!-- Text: concise and academic (no duplication with figure captions) -->
<div class="columns is-variable is-6 is-multiline">
<div class="column is-12">
<div class="content has-text-justified">
<p>
MTalk-Bench spans <em>Semantic</em>, <em>Paralinguistic</em>, and <em>Ambient</em> dimensions, with a two-level capability taxonomy
(foundational capabilities → fine-grained capabilities). Nine real-world scenarios are linked via a scenario-capability mapping to ensure ecological validity
while keeping each dialogue diagnostic by design.
</p>
<p>
We use complementary protocols: <strong>Arena-style</strong> evaluation (blind pairwise preferences aggregated as Elo for relative ordering)
and <strong>Rubric-based</strong> evaluation (hierarchical, criteria-based scoring for absolute diagnostics). Both human raters and LLM-as-a-judge
are employed to do the evaluation.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="demos">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Audio Demos</h2>
<div class="columns is-multiline is-variable is-4">
<!-- Semantic -->
<div class="column is-12-tablet is-4-desktop">
<div class="box demo-card" data-demo="Semantic">
<h3 class="title is-5">Semantic</h3>
<!-- Set 1: Comprehension & Memory -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 1</p>
<span class="tag subcap">Comprehension & Memory</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Semantic 1-1 — turn 1">
<source src="static/audio/1-1-turn1_sem.wav" type="audio/wav">
Your browser does not support the audio element.
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Semantic 1-1 — turn 2">
<source src="static/audio/1-1-turn2_sem.wav" type="audio/wav">
</audio>
</div>
</div>
<!-- Set 2: Reasoning & Execution -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 2</p>
<span class="tag subcap">Reasoning & Execution</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Semantic 1-2 — turn 1">
<source src="static/audio/1-4-turn1_sem.wav" type="audio/wav">
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Semantic 1-2 — turn 2">
<source src="static/audio/1-4-turn2_sem.wav" type="audio/wav">
</audio>
</div>
</div>
<!-- Set 3: Security & Assessment -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 3</p>
<span class="tag subcap">Security & Assessment</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Semantic 1-3 — turn 1">
<source src="static/audio/1-10-turn1_sem.wav" type="audio/wav">
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Semantic 1-3 — turn 2">
<source src="static/audio/1-10-turn2_sem.wav" type="audio/wav">
</audio>
</div>
</div>
</div>
</div>
<!-- Paralinguistic -->
<div class="column is-12-tablet is-4-desktop">
<div class="box demo-card" data-demo="Paralinguistic">
<h3 class="title is-5">Paralinguistic</h3>
<!-- Set 1: Emotion Recognition -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 1</p>
<span class="tag subcap">Emotion Recognition</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Paralinguistic 1-1 — turn 1">
<source src="static/audio/1-1-turn1_para.wav" type="audio/wav">
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Paralinguistic 1-1 — turn 2">
<source src="static/audio/1-1-turn2_para.wav" type="audio/wav">
</audio>
</div>
</div>
<!-- Set 2: Personalized Expressive Modeling -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 2</p>
<span class="tag subcap">Personalized Modeling</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Paralinguistic 1-2 — turn 1">
<source src="static/audio/1-4-turn1_para.wav" type="audio/wav">
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Paralinguistic 1-2 — turn 2">
<source src="static/audio/1-4-turn2_para.wav" type="audio/wav">
</audio>
</div>
</div>
<!-- Set 3: Feature Generation -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 3</p>
<span class="tag subcap">Paralinguistic Feature Generation</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Paralinguistic 1-3 — turn 1">
<source src="static/audio/1-10-turn1_para.wav" type="audio/wav">
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Paralinguistic 1-3 — turn 2">
<source src="static/audio/1-10-turn2_para.wav" type="audio/wav">
</audio>
</div>
</div>
</div>
</div>
<!-- Ambient -->
<div class="column is-12-tablet is-4-desktop">
<div class="box demo-card" data-demo="Ambient">
<h3 class="title is-5">Ambient</h3>
<!-- Set 1: Ambient Sound Understanding -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 1</p>
<span class="tag subcap">Ambient Sound Understanding</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Ambient sample1 — turn 1">
<source src="static/audio/1-1-turn1_amb.wav" type="audio/wav">
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Ambient sample1 — turn 2">
<source src="static/audio/1-1-turn2_amb.wav" type="audio/wav">
</audio>
</div>
</div>
<!-- Set 2: Ambient Understanding -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 2</p>
<span class="tag subcap">Ambient Sound Understanding</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Ambient sample2 — turn 1">
<source src="static/audio/1-4-turn1_amb.wav" type="audio/wav">
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Ambient sample2 — turn 2">
<source src="static/audio/1-4-turn2_amb.wav" type="audio/wav">
</audio>
</div>
</div>
<!-- Set 3: Multi-party Understanding -->
<div class="set-block">
<div class="set-header">
<p class="set-title">sample 3</p>
<span class="tag subcap">Multi-party Understanding</span>
</div>
<div class="audio-grid">
<audio class="demo-audio" controls preload="none" aria-label="Ambient sample3 — turn 1">
<source src="static/audio/1-10-turn1_amb.wav" type="audio/wav">
</audio>
<audio class="demo-audio" controls preload="none" aria-label="Ambient sample3 — turn 2">
<source src="static/audio/1-10-turn2_amb.wav" type="audio/wav">
</audio>
</div>
</div>
</div>
</div>
</div><!-- /columns -->
</div>
</section>
<section class="hero is-small" id="pdf-results">
<div class="hero-body">
<h2 class="title is-3 has-text-centered">Results Analysis</h2>
<div class="container">
<div class="carousel-container results-carousel-container">
<div class="carousel-wrapper" id="results-carousel-wrapper">
<!-- Slide: Overall results -->
<div class="carousel-item results-carousel-item">
<div class="card" style="border-radius:16px;">
<div class="card-content">
<figure class="image">
<img src="static/images/audio_eval.png" alt="Overall evaluation results">
<figcaption class="figure-caption">
<strong>Figure 3. Evaluation Results (Table 2)</strong> — Arena: Elo ratings per dimension and overall; Rubrics: absolute scores (×100). Models sorted by human overall score; ↑ = better.
</figcaption>
</figure>
</div>
</div>
</div>
<!-- Slide: Win rates across evaluators -->
<div class="carousel-item results-carousel-item">
<div class="card" style="border-radius:16px;">
<div class="card-content">
<figure class="image">
<img src="static/images/win_rates.png" alt="Win rates across evaluators">
<figcaption class="figure-caption">
<strong>Figure 4. Win rate of models</strong> — Illustrate win rates of different models across different evaluators in arena-style evaluation, revealing numerous statistically tied outcomes rather than a single dominant system.
</figcaption>
</figure>
</div>
</div>
</div>
<!-- Slide: Turn-level trends & duration correlation -->
<div class="carousel-item results-carousel-item">
<div class="card" style="border-radius:16px;">
<div class="card-content">
<figure class="image">
<img src="static/images/turn_level_trends.png" alt="Turn-level trends and duration correlation">
<figcaption class="figure-caption">
<strong>Figure 5. Turn-level trends</strong> — Quality often dips in early turns and recovers as models produce longer outputs; audio duration correlates with higher Arena/Rubrics scores, but after a minimal length the gains plateau, indicating that verbosity ≠ quality.
</figcaption>
</figure>
</div>
</div>
</div>
<!-- Slide: Bias analysis -->
<div class="carousel-item results-carousel-item">
<div class="card" style="border-radius:16px;">
<div class="card-content">
<figure class="image">
<img src="static/images/bias_analysis.png" alt="Bias analysis of evaluators">
<figcaption class="figure-caption">
<strong>Figure 6. Bias analysis of evaluators</strong> — Compared to humans (near-neutral positional preference), LLM judges exhibit measurable <em>position</em> and stronger <em>length</em> biases when scoring Arena pairs (e.g., length bias significantly positive) these structural biases should be accounted for in automated judging.
</figcaption>
</figure>
</div>
</div>
</div>
<!-- Slide: Arena–Rubrics consistency & pitfalls -->
<div class="carousel-item results-carousel-item">
<div class="card" style="border-radius:16px;">
<div class="card-content">
<figure class="image">
<img src="static/images/arena_rubrics_consistency.png" alt="Consistency and pitfalls">
<figcaption class="figure-caption">
<strong>Figure 7. Consistency of Arena and Rubrics</strong> — Internal logical consistency of each evaluator across the Arena-style and Rubric-based formats. The results show that all evaluators are validated for their judgments’ reliability regardless of the evaluation format.
</figcaption>
</figure>
</div>
</div>
</div>
</div>
<!-- Navigation -->
<button class="carousel-nav carousel-nav-left" id="results-prev" aria-label="Previous result">
<i class="fas fa-chevron-left" aria-hidden="true"></i>
</button>
<button class="carousel-nav carousel-nav-right" id="results-next" aria-label="Next result">
<i class="fas fa-chevron-right" aria-hidden="true"></i>
</button>
</div>
<!-- Pagination -->
<div class="carousel-pagination" id="results-pagination" aria-label="Results pagination">
<span class="carousel-dot active" data-slide="0"></span>
<span class="carousel-dot" data-slide="1"></span>
<span class="carousel-dot" data-slide="2"></span>
<span class="carousel-dot" data-slide="3"></span>
<span class="carousel-dot" data-slide="4"></span>
</div>
</div>
</div>
</section>
<!-- Results: Model Performance -->
<section class="section" id="results-analysis">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Experiment Results — Model Performance</h2>
<div class="content has-text-justified">
<ul class="key-findings">
<li><strong>Dimension gap:</strong> Top models have not yet reached high proficiency. They are overall strong in <em>semantics</em> but limited in <em>some specific semantic capability</em> like safety reasoning and <em>auditory cues (paralinguistic and ambient sound perception)</em>.</li>
<li><strong>Dialogue dynamics:</strong> Models recover from an early-turn quality dip by producing longer responses, trading efficiency for coherence. After a minimal length for clarity, more tokens often mean more fluff, not better answers.</li>
<li><strong>Design matters:</strong> When paired with large capacity, task specific designs deliver bigger gains than scale alone.</li>
</ul>
</div>
</div>
</section>
<!-- Meta-Evaluation: Methods & Reliability -->
<section class="section" id="meta-eval">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Meta-Evaluation — Methods & Reliability</h2>
<div class="content has-text-justified">
<ul class="key-findings">
<li><strong>Arena × Rubrics:</strong> Broadly consistent and complementary — relative ordering vs. absolute diagnostics.</li>
<li><strong>Small margins:</strong> When gaps are small, outcomes remain unstable even with more pairwise comparisons; robust conclusions appear with large gaps.</li>
<li><strong>LLM-as-judge:</strong> Agreement with humans improves when differences are clear or criteria are explicit, but LLMs show position and length biases; for nonverbal acoustic cues, reliability improves with textual annotations.</li>
<li><strong>Reliability:</strong> Good internal consistency across protocols; rubric-based assessments show stable correlations under ablations/bootstraps.</li>
</ul>
</div>
</div>
</section>
<!-- BibTeX -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@misc{2025mtalkbench,
title={MTalk-Bench: Evaluating Speech-to-Speech Models in Multi-Turn Dialogues via Arena-style and Rubrics Protocols},
author={Yuhao Du and Qianwei Huang and Guo Zhu and Zhanchen Dai and Shunian Chen and Qiming Zhu and Yuhao Zhang and Li Zhou and Benyou Wang and Haizhou Li},
year={2025},
eprint={2508.18240},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2508.18240}
}</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
<script>
// Custom Carousel Implementation
class CustomCarousel {
constructor(wrapperId, paginationId, prevBtnId, nextBtnId) {
this.wrapper = document.getElementById(wrapperId);
this.pagination = document.getElementById(paginationId);
this.prevBtn = document.getElementById(prevBtnId);
this.nextBtn = document.getElementById(nextBtnId);
this.items = this.wrapper.querySelectorAll('.carousel-item');
this.currentIndex = 0;
this.autoplayInterval = null;
this.init();
}
init() {
this.updateCarousel();
this.bindEvents();
this.startAutoplay();
}
bindEvents() {
this.prevBtn.addEventListener('click', () => this.prev());
this.nextBtn.addEventListener('click', () => this.next());
this.pagination.querySelectorAll('.carousel-dot').forEach((dot, index) => {
dot.addEventListener('click', () => this.goToSlide(index));
});
// Pause autoplay on hover
this.wrapper.addEventListener('mouseenter', () => this.stopAutoplay());
this.wrapper.addEventListener('mouseleave', () => this.startAutoplay());
}
updateCarousel() {
const translateX = -this.currentIndex * 100;
this.wrapper.style.transform = `translateX(${translateX}%)`;
// Update pagination
this.pagination.querySelectorAll('.carousel-dot').forEach((dot, index) => {
dot.classList.toggle('active', index === this.currentIndex);
});
}
next() {
this.currentIndex = (this.currentIndex + 1) % this.items.length;
this.updateCarousel();
}
prev() {
this.currentIndex = (this.currentIndex - 1 + this.items.length) % this.items.length;
this.updateCarousel();
}
goToSlide(index) {
this.currentIndex = index;
this.updateCarousel();
}
startAutoplay() {
this.stopAutoplay();
this.autoplayInterval = setInterval(() => this.next(), 5000);
}
stopAutoplay() {
if (this.autoplayInterval) {
clearInterval(this.autoplayInterval);
this.autoplayInterval = null;
}
}
}
// Initialize carousels when DOM is loaded
document.addEventListener('DOMContentLoaded', () => {
// Initialize framework figures carousel
new CustomCarousel(
'fig-carousel-wrapper',
'fig-pagination',
'fig-prev',
'fig-next'
);
// Initialize results carousel (no autoplay)
const resultsCarousel = new CustomCarousel(
'results-carousel-wrapper',
'results-pagination',
'results-prev',
'results-next'
);
resultsCarousel.stopAutoplay(); // Disable autoplay for results
});
</script>
</body>
</html>