-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
601 lines (601 loc) · 37.1 KB
/
index.html
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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width,initial-scale=1" />
<title>Speech Enhancement System Compression Using
Parameter Tying and Mixed Precision Quantization</title>
<style type="text/css">
<!-- body {-->
<!-- margin: 0 auto；-->
<!-- padding:10px 30px;-->
<!-- background: #fff;-->
<!-- color: #111;-->
<!-- font-size: 15px;-->
<!-- font-family: "Times New Roman", serif;-->
<!-- font-weight: 400;-->
<!-- line-height: 1.8;-->
<!-- -webkit-font-smoothing: antialiased;-->
<!-- }-->
audio {
width: 75%;
height:30px;
}
img {
width:100%;
}
video {
width:75%;
}
html{
height:100%;
}
body{
margin: 0 auto；
padding:10px 30px;
background: #fff;
color: #111;
height:100%;
font-size: 17px;
font-family: "Times New Roman", serif;
font-weight: 400;
line-height: 1.8;
overflow-x: hidden;
-webkit-font-smoothing: antialiased;
}
</style>
</head>
<div style="border: none; width:80%; margin: 0 auto;">
<body>
<h2 align="center">Speech Enhancement System Compression Using
Parameter Tying and Mixed Precision Quantization</h2>
<p style="line-height:100%" align="center"><b>Authors:Tao Zhong, Guinan Li, Mengzhe Geng, Junhao Xu, Zengrui Jin, Jiajun Deng, Zhaoqing Li, Tianzi Wang,
Xunying Liu</b> </p>
<p style="line-height:100%" align="justify"> <b>Abstract: </b> Speech enhancement systems are crucial components in many speech processing tasks, such as automatic speech recognition (ASR). Their practical deployment on devices requires their memory footprints and computation costs to be aggressively reduced. To this end, a novel KL divergence based parameter sharing approach is proposed for state-of-the-art temporal convolutional networks (TCNs) based audio-visual multi-channel speech separation systems featuring mask-based MVDR beamforming. Experiments are conducted on two tasks using overlapped speech simulated from the English Oxford LRS2 and Cantonese CI-AVSR datasets. The efficacy of the proposed TCN parameter tying approach is extensively shown when used as a standalone model architecture compression technique in comparison to low-rank factorization and penalized neural architecture search, as well as when combined with low-rank factorization and mixed precision low-bit precision quantization approaches. An overall “lossless” model compression ratio of up to 18.54 times is obtained over the baseline uncompressed audio-visual multi-channel speech separation system while incurring no statistically significant ASR word or character error rate increase. Comparable speech enhancement performance measured using scale-invariant signal-to-noise ratio (SI-SNR) scores is also produced. The same “lossless” speech enhancement system compression ratio is retained after applying 5-bit mixed precision quantization to the Conformer ASR back-end.</p>
<a href="#sectionI">I. Audio-visual Multi-channel Mask-based MVDR Speech Separation Front-end</a><br>
<a href="#sectionII">II. TCN Model Architecture Compression</a><br>
<a href="#sectionIII">III. Low-bit Precision Quantizatoin of TCN</a><br>
<a href="#sectionIV">IV. Experimental Setup and Results</a><br>
<a href="#sectionV">V. Reference</a>
<h3> <a name="sectionI"> I. Audio-visual Multi-channel Mask-based MVDR Speech Separation Front-end</a> </h3>
<div style="border: none; width:60%; margin: 0 auto;">
<img border="0" src="picture/neural_network.png" alt="architectures of speech enhancement front-end" height="10%">
</div>
<ul>
<li> Audio-visual multi-channel mask-based MVDR based speech separation system [1]. The TCN blocks are highlighted in blue (one in the Audio Block; three in the Target Speech and Noise Blocks each), which account for more than 84% of the overall model parameters. The speech separation front-end is on the top (the light yellow box), and the Conformer recognition back-end is on the bottom (the light red box).</li>
</ul>
<h3> <a name="sectionII"> II. TCN Model Architecture Compression</a> </h3>
<div style="border: none; width:60%; margin: 0 auto;">
<img border="0" src="picture/conv_block.png" alt="architectures of speech enhancement front-end" style="width:300px;">
</div>
<ul>
<li> Temporal convolutional network (TCN) block. Each of the 8 dilated 1-D ConvBlocks consists of two 1×1 convolutional layers located at both ends (“1x1 Conv1” and “1x1 Conv2” in blue). A depth-wise separable convolution layer (D-Conv) is located in the center. PReLU and batch normalization layers are added before and after the D-Conv layer.</li>
</ul>
<div style="border: none; width:60%; margin: 0 auto;">
<img border="0" src="picture/low_rank.png" alt="low-rank factorization" style="width:300px;">
</div>
<ul>
<li> Low-rank factorization of the two TCN 1x1 convolutional layers into four “1x1 Conv1 (left/right)” and “1x1 Conv2 (left/right)” weight matrices in blue.</li>
</ul>
<div style="border: none; width:60%; margin: 0 auto;">
<img border="0" src="picture/NAS.png" alt="penalized neural architecture search" style="width:300px;">
</div>
<ul>
<li> Penalized neural architecture search over the projection layer dimensionality for low-rank factorized TCN 1x1convolutional layers ranging from 40 to 170.</li>
</ul>
<div style="border: none; width:60%; margin: 0 auto;">
<img border="0" src="picture/share.png" alt="parameter sharing" style="width:300px;">
</div>
<ul>
<li> Parameter sharing among low-rank factorized weight matrices located at comparable positions (highlighted in the same yellow boxes) across different dilated 1-D ConvBlocks within a TCN.</li>
</ul>
<h3> <a name="sectionIII"> III. Low-bit Precision Quantizatoin of TCN</a> </h3>
<div style="border: none; width:60%; margin: 0 auto;">
<img border="0" src="picture/kl.png" alt="KL divergence based mixed precision quantization" style="width:300px;">
</div>
<ul>
<li> Mixed precision quantization of low-rank factorized TCN’s 1x1 convolutional layers (shown "1x1 Conv1 (left/right)" and "1x1 Conv2 (left/right)") with their layer level varying precision settings learned using KL divergence [2].</li>
</ul>
<div style="border: none; width:60%; margin: 0 auto;">
<img border="0" src="picture/hessian.png" alt="Hessian trace based mixed precision quantization" style="width:300px;">
</div>
<ul>
<li> Mixed precision quantization of low-rank factorized TCN’s 1x1 convolutional layers (shown "1x1 Conv1 (left/right)" and "1x1 Conv2 (left/right)") with their layer level varying precision settings learned using Hessian trace.</li>
</ul>
<div style="border: none; width:60%; margin: 0 auto;">
<img border="0" src="picture/nas_q.png" alt="penalized NAS based mixed precision quantization" style="width:300px;">
</div>
<ul>
<li> Mixed precision quantization of low-rank factorized TCN’s 1x1 convolutional layers (shown "1x1 Conv1 (left/right)" and "1x1 Conv2 (left/right)") with their layer level varying precision settings learned using penalized NAS.</li>
</ul>
<h3> <a name="sectionIV">IV. Experimental Setup and Results</a></h3>
<div style="border: none; width:100%; margin: 0 auto;">
<h4> A. Experimental setup</h4>
<h4> (1) Oxford LRS2 data simulated overlapped speech </h4>
<ul>
<li> A 15-channel symmetric linear array with non-even inter-channel spacing is leveraged to simulate the multi-channel overlapped speech using the Oxford LRS2 dataset [3] with 96997, 4272 and 4972 utterances respectively for training (91.37 hours), validation (2.59 hours) and test (2.32 hours). </li>
<li> Espnet style Conformer AED ASR model contains 12 encoder and 6 decoder layers is used in the submitted paper. </li>
<li> In all the experiments of the Oxford LRS2 data simulated overlapped speech, “lossless” compression is achieved when no statistically significant (MAPSSWE [4], α= 0.05) WER increase is observed after model compression. </li>
</ul>
<h4> (2) CI-AVSR simulated overlapped speech </h4>
<ul>
<li> A 3-channel symmetric linear array with even inter-channel spacing is leveraged to simulate the multi-channel overlapped speech using the Cantonese CI-AVSR in-car command speech corpus [5] with 89804, 12844 and 13182 utterances respectively for training (68.8 hours), validation (9.4 hours) and test (9.4 hours). </li>
<li> Espnet style Conformer AED ASR model contains 12 encoder and 6 decoder layers is used in the submitted paper. </li>
<li> In all the experiments of the CI-AVSR audio-visual Cantonese in-car speech simulated overlapped speech, “lossless” compression is achieved when no statistically significant (MAPSSWE [4], α= 0.05) CER increase is observed after model compression. </li>
</ul>
<h4> B. Experimental results on Oxford LRS2 data simulated overlapped speech </h4>
<table border="1">
<tr>
<th rowspan="2" colspan="2"></th>
<th rowspan="2" colspan="2">Maximum “lossless” compression ratio</th>
<th colspan="4"> Four ranges of the angle difference between the target and interfering speakers relative to the microphone array </th>
</tr>
<tr>
<th>[0°-15°)</th>
<th>[15°-45°)</th>
<th>[45°-90°)</th>
<th>[90°-180°)</th>
</tr>
<tr>
<th colspan="2"> Target speaker video</th>
<th colspan="2"> - </th>
<td style="text-align:center">
<video src="test_audio/clean/transm5-6331559613336179781-00019.mp4" controls="controls">
your browser does not support the video tag
</video>
</td>
<td style="text-align:center">
<video src="test_audio/clean/transm5-6331559613336179781-00027.mp4" controls="controls">
your browser does not support the video tag
</video>
</td>
<td style="text-align:center">
<video src=test_audio/clean/transm5-6331559613336179781-00029.mp4 controls="controls">
your browser does not support the video tag
</video>
</td>
<td style="text-align:center">
<video src=test_audio/clean/transm5-6332062124509813446-00018.mp4 controls="controls">
your browser does not support the video tag
</video>
</td>
</tr>
<tr>
<th colspan="2"> Overlapped speech</th>
<th colspan="2"> - </th>
<td style="text-align:center">
<img src="test_audio/simulated/0-15/0-mc/mixture/00019.png" alt>
<audio src="test_audio/simulated/0-15/0-mc/mixture/00019.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/simulated/15-45/0-mc/mixture/00027.png" alt>
<audio src="test_audio/simulated/15-45/0-mc/mixture/00027.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/simulated/45-90/0-mc/mixture/00029.png" alt>
<audio src="test_audio/simulated/45-90/0-mc/mixture/00029.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/simulated/90-180/0-mc/mixture/00018.png" alt>
<audio src="test_audio/simulated/90-180/0-mc/mixture/00018.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Clean speech</th>
<th colspan="2">-</th>
<td style="text-align:center">
<img src="test_audio/simulated/0-15/0-mc/clean/00019.png" alt>
<audio src="test_audio/simulated/0-15/0-mc/clean/00019.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/simulated/15-45/0-mc/clean/00027.png" alt>
<audio src="test_audio/simulated/15-45/0-mc/clean/00027.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/simulated/45-90/0-mc/clean/00029.png" alt>
<audio src="test_audio/simulated/45-90/0-mc/clean/00029.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/simulated/90-180/0-mc/clean/00018.png" alt>
<audio src="test_audio/simulated/90-180/0-mc/clean/00018.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR <br> (uncompressed model) <br> (Sys. 1 in TABLE IV in the submitted paper) </th>
<th colspan="2">1.0 </th>
<td style="text-align:center">
<img src="test_audio/baseline/rev1-6331559613336179781-00019.png" alt >
<audio src="test_audio/baseline/rev1-6331559613336179781-00019.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/baseline/rev2-6331559613336179781-00027.png" alt >
<audio src="test_audio/baseline/rev2-6331559613336179781-00027.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/baseline/rev3-6331559613336179781-00029.png" alt >
<audio src="test_audio/baseline/rev3-6331559613336179781-00029.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/baseline/rev4-6332062124509813446-00018.png" alt >
<audio src="test_audio/baseline/rev4-6332062124509813446-00018.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR + <br> Low Rank <br> (Sys. 2 in TABLE IV in the submitted paper)</th>
<th colspan="2">2.07</th>
<td style="text-align:center">
<img src="test_audio/low_rank/rev1-6331559613336179781-00019.png" alt >
<audio src="test_audio/low_rank/rev1-6331559613336179781-00019.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank/rev2-6331559613336179781-00027.png" alt >
<audio src="test_audio/low_rank/rev2-6331559613336179781-00027.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank/rev3-6331559613336179781-00029.png" alt >
<audio src="test_audio/low_rank/rev3-6331559613336179781-00029.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank/rev4-6332062124509813446-00018.png" alt >
<audio src="test_audio/low_rank/rev4-6332062124509813446-00018.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR + <br>Low Rank + <br> Penalized NAS <br> (Sys. 4 in TABLE IV in the submitted paper)</th>
<th colspan="2">2.87</th>
<td style="text-align:center">
<img src="test_audio/low_rank+nas/rev1-6331559613336179781-00019.png" alt >
<audio src="test_audio/low_rank+nas/rev1-6331559613336179781-00019.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+nas/rev2-6331559613336179781-00027.png" alt >
<audio src="test_audio/low_rank+nas/rev2-6331559613336179781-00027.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+nas/rev3-6331559613336179781-00029.png" alt >
<audio src="test_audio/low_rank+nas/rev3-6331559613336179781-00029.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+nas/rev4-6332062124509813446-00018.png" alt >
<audio src="test_audio/low_rank+nas/rev4-6332062124509813446-00018.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR + <br>Low Rank + <br> Parameter Sharing <br> (Sys. 5 in TABLE IV in the submitted paper)</th>
<th colspan="2">3.93</th>
<td style="text-align:center">
<img src="test_audio/low_rank+parameter_sharing/rev1-6331559613336179781-00019.png" alt >
<audio src="test_audio/low_rank+parameter_sharing/rev1-6331559613336179781-00019.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+parameter_sharing/rev2-6331559613336179781-00027.png" alt >
<audio src="test_audio/low_rank+parameter_sharing/rev2-6331559613336179781-00027.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+parameter_sharing/rev3-6331559613336179781-00029.png" alt >
<audio src="test_audio/low_rank+parameter_sharing/rev3-6331559613336179781-00029.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+parameter_sharing/rev4-6332062124509813446-00018.png" alt >
<audio src="test_audio/low_rank+parameter_sharing/rev4-6332062124509813446-00018.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR + <br>Low Rank + <br> Parameter Sharing + <br> KL Divergence Based <br> Mixed Precision Quantization <br> (Sys. 16 in TABLE IV in the submitted paper)</th>
<th colspan="2">18.54</th>
<td style="text-align:center">
<img src="test_audio/low_rank+parameter_sharing+kl_based_quantization/rev1-6331559613336179781-00019.png" alt >
<audio src="test_audio/low_rank+parameter_sharing+kl_based_quantization/rev1-6331559613336179781-00019.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+parameter_sharing+kl_based_quantization/rev2-6331559613336179781-00027.png" alt >
<audio src="test_audio/low_rank+parameter_sharing+kl_based_quantization/rev2-6331559613336179781-00027.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+parameter_sharing+kl_based_quantization/rev3-6331559613336179781-00029.png" alt >
<audio src="test_audio/low_rank+parameter_sharing+kl_based_quantization/rev3-6331559613336179781-00029.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio/low_rank+parameter_sharing+kl_based_quantization/rev4-6332062124509813446-00018.png" alt >
<audio src="test_audio/low_rank+parameter_sharing+kl_based_quantization/rev4-6332062124509813446-00018.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
</table>
<h4> C. Experimental results on CI-AVSR audio-visual Cantonese in-car speech simulated overlapped speech </h4>
<table border="1">
<tr>
<th rowspan="2" colspan="2"></th>
<th rowspan="2" colspan="2">Maximum “lossless” compression ratio</th>
<th rowspan="2" colspan="4"> Examples </th>
</tr>
<tr>
</tr>
<tr>
<th colspan="2"> Target speaker video</th>
<th colspan="2"> - </th>
<td style="text-align:center">
<video src="test_audio_ciavsr/clean/modify/1_40.mp4" controls="controls">
your browser does not support the video tag
</video>
</td>
<td style="text-align:center">
<video src="test_audio_ciavsr/clean/modify/2_90.mp4" controls="controls">
your browser does not support the video tag
</video>
</td>
<td style="text-align:center">
<video src=test_audio_ciavsr/clean/modify/2_98.mp4 controls="controls">
your browser does not support the video tag
</video>
</td>
<td style="text-align:center">
<video src=test_audio_ciavsr/clean/modify/2_7.mp4 controls="controls">
your browser does not support the video tag
</video>
</td>
</tr>
<tr>
<th colspan="2"> Overlapped speech</th>
<th colspan="2"> - </th>
<td style="text-align:center">
<img src="test_audio_ciavsr/mix/0/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt>
<audio src="test_audio_ciavsr/mix/0/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/mix/0/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt>
<audio src="test_audio_ciavsr/mix/0/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/mix/0/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt>
<audio src="test_audio_ciavsr/mix/0/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/mix/0/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.png" alt>
<audio src="test_audio_ciavsr/mix/0/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Clean speech</th>
<th colspan="2">-</th>
<td style="text-align:center">
<img src="test_audio_ciavsr/official_clean/15/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt>
<audio src="test_audio_ciavsr/official_clean/15/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/official_clean/15/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt>
<audio src="test_audio_ciavsr/official_clean/15/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/official_clean/15/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt>
<audio src="test_audio_ciavsr/official_clean/15/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/official_clean/15/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.png" alt>
<audio src="test_audio_ciavsr/official_clean/15/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR <br> (uncompressed model) <br> (Sys. 1 in TABLE VI in the submitted paper) </th>
<th colspan="2">1.0 </th>
<td style="text-align:center">
<img src="test_audio_ciavsr/baseline/final_select/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/baseline/final_select/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/baseline/final_select/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/baseline/final_select/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/baseline/final_select/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/baseline/final_select/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/baseline/final_select/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.png" alt >
<audio src="test_audio_ciavsr/baseline/final_select/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR + <br> Low Rank <br> (Sys. 2 in TABLE VI in the submitted paper)</th>
<th colspan="2">1.41</th>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank/final_select/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank/final_select/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank/final_select/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank/final_select/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank/final_select/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank/final_select/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank/final_select/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.png" alt >
<audio src="test_audio_ciavsr/low_rank/final_select/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR + <br>Low Rank + <br> Penalized NAS <br> (Sys. 4 in TABLE VI in the submitted paper)</th>
<th colspan="2">1.70</th>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+nas/final_select/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+nas/final_select/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+nas/final_select/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+nas/final_select/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+nas/final_select/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+nas/final_select/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+nas/final_select/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+nas/final_select/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR + <br>Low Rank + <br> Parameter Sharing <br> (Sys. 5 in TABLE VI in the submitted paper)</th>
<th colspan="2">2.71</th>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+parameter_sharing/final_select/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+parameter_sharing/final_select/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+parameter_sharing/final_select/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+parameter_sharing/final_select/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+parameter_sharing/final_select/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+parameter_sharing/final_select/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+parameter_sharing/final_select/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+parameter_sharing/final_select/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
<tr>
<th colspan="2">Audio-visual Mask-based MVDR + <br>Low Rank + <br> Parameter Sharing + <br> KL Divergence Based <br> Mixed Precision Quantization <br> (Sys. 16 in TABLE VI in the submitted paper)</th>
<th colspan="2">13.84</th>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+parameter_sharing_kl_based_quantization/final_selected/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+parameter_sharing_kl_based_quantization/final_selected/C010-ICAVSP-00009-HKUSTXXX-XXXFXX-ih_10000040-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+parameter_sharing_kl_based_quantization/final_selected/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+parameter_sharing_kl_based_quantization/final_selected/C010-ICAVSP-00015-HKUSTXXX-XXXMXX-ih_20000090-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+parameter_sharing_kl_based_quantization/final_selected/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+parameter_sharing_kl_based_quantization/final_selected/C010-ICAVSP-00023-HKUSTXXX-XXXFXX-ih_20000098-2spk_S1_S3-WindscreenWiper_Horn_RoadAmbiance_Alarm_CarDoor-SameGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
<td style="text-align:center">
<img src="test_audio_ciavsr/low_rank+parameter_sharing_kl_based_quantization/final_selected/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.png" alt >
<audio src="test_audio_ciavsr/low_rank+parameter_sharing_kl_based_quantization/final_selected/C011-ICAVSP-00007-HKUSTXXX-XXXMXX-ih_20000007-2spk_S1_S4-BackgroundMusic_Rain_Hail_Ignition-DifferentGender.wav" controls="controls">
Your browser does not support the audio element.
</audio>
</td>
</tr>
</table>
</div>
<h3><a name="sectionV">V. Reference</a></h3>
<p> [1] J. Yu et al., “Audio-Visual Multi-Channel Integration and Recognition of Overlapped Speech,” IEEE T AUDIO SPEECH, 2021.</p>
<p> [2] J. Xu et al., “Mixed Precision DNN Quantization for Overlapped Speech Separation and Recognition,” in ICASSP, 2022.</p>
<p> [3] J. Son Chung et al., “Lip Reading Sentences in the Wild,” in CVPR, 2017.</p>
<p> [4] L. Gillick et al., “Some statistical issues in the comparison of speech recognition algorithms,” in ICASSP, 1989.</p>
<p> [5] W. Dai et al., “CI-AVSR: A Cantonese Audio-Visual Speech Datasetfor In-car Command Recognition,” in LREC, 2022.</p>
</body>
</div>
</html>