-
Notifications
You must be signed in to change notification settings - Fork 1
/
index.html
1712 lines (1584 loc) · 59.4 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
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>
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<title>Deep learning for music separation</title>
<link rel="stylesheet" href="css/reveal.css">
<link rel="stylesheet" href="css/theme/lirmm.css">
<link href="fontawesome-free-5.3.1-web/css/all.css" rel="stylesheet">
<!-- <link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.1.0/css/all.css" integrity="sha384-lKuwvrZot6UHsBSfcMvOkWwlCMgc0TaWr+30HWe3a4ltaBwTZhyTEggF5tJv8tbt" crossorigin="anonymous"> -->
<!-- Theme used for syntax highlighting of code -->
<link rel="stylesheet" href="lib/css/github.css">
<!-- Printing and PDF exports -->
<script>
var link = document.createElement( 'link' );
link.rel = 'stylesheet';
link.type = 'text/css';
link.href = window.location.search.match( /print-pdf/gi ) ? 'css/print/pdf.css' : 'css/print/paper.css';
document.getElementsByTagName( 'head' )[0].appendChild( link );
</script>
</head>
<body>
<div class="reveal">
<div class="slides">
<section id="cover" data-background="css/theme/img/background-red.jpg" data-background-repeat="no-repeat" data-background-size="cover" data-background-position="0 0" data-state="no-title-footer no-progressbar has-dark-background">
<h2 id="covertitle">Deep learning for music separation</h2>
<p id="coverauthors">
Antoine Liutkus & Fabian-Robert Stöter<br/>
Inria and LIRMM, Montpellier<br/>
<p>
[email protected]<br />
</p>
<p>
[email protected]<br />
<i class="fab fa-twitter"></i>faroit
</p>
</p>
<p>
September 2nd, 2019
</p>
<p>
<img src="css/theme/img/inria-cover.svg" id="inria" class="logo" alt="">
</section>
<!-- Introduction
In an introductory section, we will motivate the tutorial by explaining how music separation with DNN emerged with
data-driven methods coming from machine-learning or image processing communities. This comes with machine-learning
tricks to make methods work in practice. Meanwhile, many audio processing good practices are often forgotten or not
correctly applied, although they are mandatory for good performance.
• Timeline for DNN music separation
A brief history of the topic in the bigger picture: machine learning on the one hand, source separation on the other
hand.
• Outline
The baseline system: what is the state of the art? Quick literature review
How to implement and make it work? The distinction between training and testing.
An audio demo of the critical impact of engineering for the “same” system.
-->
<!-- INTRODUCTION -->
<section>
<h1>Music Unmixing/Separation</h1>
<img width="45%" style="float:left" src="assets/intro1.png" alt="">
<img class="fragment" width="45%" src="assets/intro2.png" alt="">
</section>
<section>
<h1>Applications</h1>
<img width="60%" style="float:right" src="assets/karaoke.jpg" alt="">
<ul>
<li>Automatic Karaoke</li>
<li>Creative Music Production</li>
<li>Active listening</li>
<li>Upmixing (stereo $\Rightarrow$ 5.1)</li>
<li>Music Education</li>
<li>Pre-processing for MIR</li>
</ul>
</section>
<!-- <section>
<iframe class="stretch" data-src="audio/player_x.html"></iframe>
</section> -->
<section>
<h1>Unmatched state of the art</h1>
<ul>
<li>Very active research community and evaluations
<ul>
<li>International campaigns: MIREX, SiSEC</li>
<li>Not-so-recent fact: <strong>separation with DNN works</strong></li>
</ul>
<p>
<img width="30%" style="float:right; margin-top:2%;" src="assets/stefanyuki.svg" alt="">
<li style="margin-top:4%;">State of the art: SONY corporation systems
<ul>
<li style="color:gray">S. Uhlich et al. "Deep neural network based instrument extraction from music." ICASSP 2015.</li>
$\Rightarrow$ Vocals SDR: 5dB (SiSEC 2016)
<li style="color:gray">S. Uhlich, et al. "Improving music source separation based on deep neural networks through data augmentation and network blending." ICASSP 2017.</li>
$\Rightarrow$ Vocals SDR: 5.9dB (SiSEC 2018)
</li></ul>
<p>
<li style="margin-top:4%;">Open (and popular) implementations 2.5 dB behind state of the art!
<ul>
<li><a href=https://github.com/MTG/DeepConvSep><i class="fab fa-github"></i>/MTG/DeepConvSep</a> (349<i class="fa fa-star" aria-hidden="true"></i>) 2.5 dB vocals SDR (SiSEC'16)
<div style="color:gray"><font size="4">P. Chandna et al. "Monoaural audio source separation using deep convolutional neural networks", LVA-ICA, 2017.</font></div>
</li>
<li><a href=https://github.com/f90/Wave-U-Net><i class="fab fa-github"></i>/f90/Wave-U-Net</a> (309<i class="fa fa-star" aria-hidden="true"></i>) 3.3 dB vocals SDR (SiSEC'18)
<div style="color:gray"><font size="4">D. Stoller "Wave-u-net: A multi-scale neural network for end-to-end audio source separation." arXiv, 2018.</font></div>
</li>
</ul>
</ul>
</section>
<section>
<h1>Motivations of this tutorial</h1>
<h2>Understand source separation</h2>
<ul>
<li>Signal processing aspects</li>
<li>Quick overview of the topic</li>
<li>Discriminative and generative methods</li>
</ul>
<p>
<h2>Understand deep neural nets</h2>
<ul>
<li>Fundamental models for static/temporal data</li>
<li>A starter on training</li>
<li>Models for audio</li>
</ul>
<p>
<h2>Python practice</h2>
<ul>
<li>How to implement and train deep nets with Pytorch</li>
<li>Official release of <code>open-unmix</code> today!
<a href=https://github.com/sigsep/open-unmix-pytorch><i class="fab fa-github"></i>/sigsep/open-unmix-pytorch</a>
</li>
<li>MIT-licensed state of the art performance</li>
</ul></li>
</ul>
<div style="margin-top:8%; color:gray">
All slides and material available at:
<a href=https://github.com/sigsep><i class="fab fa-github"></i>/sigsep</a>
</div>
</section>
<section>
<h1>What <code>open-unmix</code> can achieve</h1>
<iframe src='http://umx-sisec18.s3-website.eu-west-3.amazonaws.com/' style='height:500px;width:100%'></iframe>
</section>
<section>
<h1>Tutorial general outline</h1>
<ul>
<li>Signal processing basics</li>
<li>Evaluating source separation</li>
<li>Datasets</li>
<li><font color="red">Hands on oracle separation</font></li>
<p>
<li>A brief history of music separation</li>
<li>A starter on deep neural networks</li>
<li>Discriminative and generative separation</li>
<li><font color="red">Hands on using pre-trained <code>open-unmix</code></font></li>
<p>
<li>Training a DNN</li>
<li>Audio datasets</li>
<li><font color="red">Hands on training on pytorch</font></li>
<p>
<li>The <code>open-unmix</code> story</li>
<li>Testing tricks</li>
<li><font color="red">Hands on testing tricks with <code>open-unmix</code></font></li>
<p>
<li>Conclusion</li>
</ul>
</section>
<section>
<h1>Signal processing</h1>
<h2>Time-frequency representations</h2>
</section>
<section data-background-transition="none" data-state="no-title-footer" data-background-image="assets/demotrack/mix.jpg">
<!-- <h1><button data-audio="assets/demotrack/mixture.m4a">▶</button></h1> -->
<h1 style="margin-top:5%; color:white">Mixture spectrogram</h1>
</section>
<section data-background-transition="none">
<section data-background-transition="none" data-state="no-title-footer" data-background-image="assets/demotrack/vocals.jpg">
<!-- <h1><button data-audio="assets/demotrack/vocals.m4a">▶</button></h1> -->
<h1 style="margin-top:5%; color:white">Vocals spectrogram</h1>
</section>
<section data-background-transition="none" data-state="no-title-footer" data-background-image="assets/demotrack/drums.jpg">
<!-- <h1><button data-audio="assets/demotrack/drums.m4a">▶</button></h1> -->
<h1 style="margin-top:5%; color:white">Drums spectrogram</h1>
</section>
<section data-background-transition="none" data-state="no-title-footer" data-background-image="assets/demotrack/bass.jpg">
<!-- <h1><button data-audio="assets/demotrack/bass.m4a">▶</button></h1> -->
<h1 style="margin-top:5%; color:white">Bass spectrogram</h1>
</section>
</section>
<section>
<h2>Spectral analysis as pre-whitening</h2>
<img style="float:right" width="50%" src="assets/fourier_whitening.svg" alt="">
<ul>
<li>Frames too short: not diagonalized</li>
<li>Frames too long: not stationary</li>
</ul>
</section>
<section>
<!-- <h1>Time-frequency representations</h1> -->
<h2>Spectral analysis as pre-whitening</h2>
<img width="90%" src="assets/STFT.gif" alt="">
</section>
<section>
<section>
<h1>Filtering</h1>
<img width="75%" src="assets/soft_masking_1.svg" alt="">
</section>
<section transition='none'>
<h1>Filtering</h1>
<img width="75%" src="assets/soft_masking_2.svg" alt="">
</section>
<section transition='none'>
<h1>Filtering</h1>
<img width="75%" src="assets/soft_masking_3.svg" alt="">
</section>
<section transition='none'>
<h1>Filtering</h1>
<img width="75%" src="assets/soft_masking_4.svg" alt="">
</section>
<section data-transition="none">
<h1>Filtering</h1>
<img width="75%" src="assets/soft_masking_5.svg" alt="">
<img class="fragment" style="float:right" width="100%" src="assets/soft_masking_formula_potatoes.svg" alt="">
</section>
<section transition='none'>
<h1>Filtering</h1>
<img width="75%" src="assets/soft_masking_4.svg" alt="">
<img style="float:right" width="100%" src="assets/soft_masking_formula.svg" alt="">
</section>
</section>
<section>
<h1>Evaluating quality</h1>
<img style="float:right" width="100%" src="assets/evaluation_big_picture.svg" alt="">
</section>
<section>
<h1>Perceptual evaluation: hot topics</h1>
<ul>
<li>Which questions to ask ?
<p>
<h6 style="margin-top:3%; color:gray">
E. Cano et al. "The dimensions of perceptual quality of sound source separation." ICASSP, 2018.
</h6>
</li>
<li>Referenceless evaluation
<h6 style="margin-top:3%; color:gray">
E. Grais et al. "Referenceless Performance Evaluation of Audio Source Separation using Deep Neural Networks." arXiv:1811.00454 (2018)</li>
</h6>
<li>Crowdsourced evaluations
<h6 style="margin-top:3%; color:gray">
M. Cartwright et al. "Crowdsourced Pairwise-Comparison for Source Separation Evaluation." ICASSP, 2018.</li>
</h6>
</section>
<section>
<h1>Objective evaluation</h1>
<h2><img width="50px"src="assets/matlab.jpg" style="position:relative;left:0px; top:10px;" alt="">BSSeval v3</h2>
All metrics in dB. The higher, the better:
<ul>
<li><strong>SDR</strong>: Source to distortion ratio. <em>Error in the estimate</em>.</li>
<li><strong>SIR</strong>: Source to interference ratio. <em>Presence of other sources</em>.</li>
<li><strong>SAR</strong>: Source to artifacts ratio. <em>Amount of artificial noise</em>.</li>
</ul>
<p>
<h6 style="margin-top:2%; color:gray">
E. Vincent et al. "Performance measurement in blind audio source separation."
IEEE TASLP 14.4 (2006): 1462-1469.
</h6>
<h2><img width="50px"src="assets/python.jpg" style="position:relative;left:0px; top:25px;" alt="">museval (BSSeval v4)</h2>
<ul>
<li><strong>Better</strong> matching filters computed track-wise </li>
<li><strong>Faster</strong> 10x</li>
<li><code>pip install museval</code>, <a href=https://github.com/sigsep/sigsep-mus-eval><i class="fab fa-github"></i>/sigsep/sigsep-mus-eval</a></li>
</ul>
<h6 style="margin-top:2%; color:gray">
F. Stöter et al. "The 2018 Signal Separation Evaluation Campaign."
LVA/ICA 2018.
</h6>
</section>
<section>
<h1>Oracle evaluation</h1>
<img style="float:right" width="100%" src="assets/oracle_evaluation.svg" alt="">
</section>
<section>
<h1>Music separation datasets</h1>
<table>
<thead>
<tr>
<th><strong>Name </strong></th>
<th><strong>Year</strong></th>
<th><strong>Reference </strong></th>
<th><strong>#Tracks</strong></th>
<th><strong>Tracks dur (s) </strong></th>
<th><strong>Full/stereo?</strong></th>
<th><strong>Total length</strong></th>
</tr>
</thead>
<tbody>
<tr>
<td>
<a href="http://www.mtg.upf.edu/download/datasets/mass" target="_blank" rel="noopener noreferrer">
MASS
<svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" x="0px" y="0px" viewBox="0 0 100 100" width="15" height="15" class="icon outbound">
<path fill="currentColor" d="M18.8,85.1h56l0,0c2.2,0,4-1.8,4-4v-32h-8v28h-48v-48h28v-8h-32l0,0c-2.2,0-4,1.8-4,4v56C14.8,83.3,16.6,85.1,18.8,85.1z"></path>
<polygon fill="currentColor" points="45.7,48.7 51.3,54.3 77.2,28.5 77.2,37.2 85.2,37.2 85.2,14.9 62.8,14.9 62.8,22.9 71.5,22.9"></polygon>
</svg>
</a>
</td>
<td>2008</td>
<td>(Vinyes)</td>
<td>9</td>
<td>
(16
<span class="katex">
<span class="katex-mathml">
<math>
<semantics>
<mrow>
<mo>±</mo>
</mrow>
<annotation encoding="application/x-tex">\pm</annotation>
</semantics>
</math>
</span>
<span aria-hidden="true" class="katex-html"><span class="strut" style="height: 0.58333em;"></span><span class="strut bottom" style="height: 0.66666em; vertical-align: -0.08333em;"></span><span class="base"><span class="mord">±</span></span></span>
</span>
7)
</td>
<td>❌ / ✔️</td>
<td>2m24s</td>
</tr>
<tr>
<td>
<a href="https://sites.google.com/site/unvoicedsoundseparation/mir-1k" target="_blank" rel="noopener noreferrer">
MIR-1K
<svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" x="0px" y="0px" viewBox="0 0 100 100" width="15" height="15" class="icon outbound">
<path fill="currentColor" d="M18.8,85.1h56l0,0c2.2,0,4-1.8,4-4v-32h-8v28h-48v-48h28v-8h-32l0,0c-2.2,0-4,1.8-4,4v56C14.8,83.3,16.6,85.1,18.8,85.1z"></path>
<polygon fill="currentColor" points="45.7,48.7 51.3,54.3 77.2,28.5 77.2,37.2 85.2,37.2 85.2,14.9 62.8,14.9 62.8,22.9 71.5,22.9"></polygon>
</svg>
</a>
</td>
<td>2010</td>
<td>(Hsu and Jang)</td>
<td>1,000</td>
<td>8 <span class="katex"><span class="katex-mathml"></td>
<td>❌ / ❌</td>
<td>2h13m20s</td>
</tr>
<tr>
<td>
<a href="http://www.tsi.telecom-paristech.fr/aao/en/2012/03/12/quasi/" target="_blank" rel="noopener noreferrer">
QUASI
<svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" x="0px" y="0px" viewBox="0 0 100 100" width="15" height="15" class="icon outbound">
<path fill="currentColor" d="M18.8,85.1h56l0,0c2.2,0,4-1.8,4-4v-32h-8v28h-48v-48h28v-8h-32l0,0c-2.2,0-4,1.8-4,4v56C14.8,83.3,16.6,85.1,18.8,85.1z"></path>
<polygon fill="currentColor" points="45.7,48.7 51.3,54.3 77.2,28.5 77.2,37.2 85.2,37.2 85.2,14.9 62.8,14.9 62.8,22.9 71.5,22.9"></polygon>
</svg>
</a>
</td>
<td>2011</td>
<td>(Liutkus et al.)</td>
<td>5</td>
<td>
(206
<span class="katex">
<span class="katex-mathml">
<math>
<semantics>
<mrow>
<mo>±</mo>
</mrow>
<annotation encoding="application/x-tex">\pm</annotation>
</semantics>
</math>
</span>
<span aria-hidden="true" class="katex-html"><span class="strut" style="height: 0.58333em;"></span><span class="strut bottom" style="height: 0.66666em; vertical-align: -0.08333em;"></span><span class="base"><span class="mord">±</span></span></span>
</span>
21)
</td>
<td>✔️ / ✔️</td>
<td>17m10s</td>
</tr>
<tr>
<td>
<a href="http://www.loria.fr/~aliutkus/kam/" target="_blank" rel="noopener noreferrer">
ccMixter
<svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" x="0px" y="0px" viewBox="0 0 100 100" width="15" height="15" class="icon outbound">
<path fill="currentColor" d="M18.8,85.1h56l0,0c2.2,0,4-1.8,4-4v-32h-8v28h-48v-48h28v-8h-32l0,0c-2.2,0-4,1.8-4,4v56C14.8,83.3,16.6,85.1,18.8,85.1z"></path>
<polygon fill="currentColor" points="45.7,48.7 51.3,54.3 77.2,28.5 77.2,37.2 85.2,37.2 85.2,14.9 62.8,14.9 62.8,22.9 71.5,22.9"></polygon>
</svg>
</a>
</td>
<td>2014</td>
<td>(Liutkus et al)</td>
<td>50</td>
<td>
(231
<span class="katex">
<span class="katex-mathml">
<math>
<semantics>
<mrow>
<mo>±</mo>
</mrow>
<annotation encoding="application/x-tex">\pm</annotation>
</semantics>
</math>
</span>
<span aria-hidden="true" class="katex-html"><span class="strut" style="height: 0.58333em;"></span><span class="strut bottom" style="height: 0.66666em; vertical-align: -0.08333em;"></span><span class="base"><span class="mord">±</span></span></span>
</span>
77)
</td>
<td>✔️ / ✔️</td>
<td>3h12m30s</td>
</tr>
<tr>
<td>
<a href="http://medleydb.weebly.com/" target="_blank" rel="noopener noreferrer">
MedleyDB
<svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" x="0px" y="0px" viewBox="0 0 100 100" width="15" height="15" class="icon outbound">
<path fill="currentColor" d="M18.8,85.1h56l0,0c2.2,0,4-1.8,4-4v-32h-8v28h-48v-48h28v-8h-32l0,0c-2.2,0-4,1.8-4,4v56C14.8,83.3,16.6,85.1,18.8,85.1z"></path>
<polygon fill="currentColor" points="45.7,48.7 51.3,54.3 77.2,28.5 77.2,37.2 85.2,37.2 85.2,14.9 62.8,14.9 62.8,22.9 71.5,22.9"></polygon>
</svg>
</a>
</td>
<td>2014</td>
<td>(Bittner et al)</td>
<td>63</td>
<td>
(206
<span class="katex">
<span class="katex-mathml">
<math>
<semantics>
<mrow>
<mo>±</mo>
</mrow>
<annotation encoding="application/x-tex">\pm</annotation>
</semantics>
</math>
</span>
<span aria-hidden="true" class="katex-html"><span class="strut" style="height: 0.58333em;"></span><span class="strut bottom" style="height: 0.66666em; vertical-align: -0.08333em;"></span><span class="base"><span class="mord">±</span></span></span>
</span>
121)
</td>
<td>✔️ / ✔️</td>
<td>3h36m18s</td>
</tr>
<tr>
<td>
<a href="http://mac.citi.sinica.edu.tw/ikala/" target="_blank" rel="noopener noreferrer">
iKala
<svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" x="0px" y="0px" viewBox="0 0 100 100" width="15" height="15" class="icon outbound">
<path fill="currentColor" d="M18.8,85.1h56l0,0c2.2,0,4-1.8,4-4v-32h-8v28h-48v-48h28v-8h-32l0,0c-2.2,0-4,1.8-4,4v56C14.8,83.3,16.6,85.1,18.8,85.1z"></path>
<polygon fill="currentColor" points="45.7,48.7 51.3,54.3 77.2,28.5 77.2,37.2 85.2,37.2 85.2,14.9 62.8,14.9 62.8,22.9 71.5,22.9"></polygon>
</svg>
</a>
</td>
<td>2015</td>
<td>(Chan et al)</td>
<td>206</td>
<td>30</td>
<td>❌ / ❌</td>
<td>1h43m</td>
</tr>
<tr>
<td><a href="/datasets/dsd100.html" class="">DSD100</a><span class="badge"></span></td>
<td>2015</td>
<td>(Ono et al)</td>
<td>100</td>
<td>
(251
<span class="katex">
<span class="katex-mathml">
<math>
<semantics>
<mrow>
<mo>±</mo>
</mrow>
<annotation encoding="application/x-tex">\pm</annotation>
</semantics>
</math>
</span>
<span aria-hidden="true" class="katex-html"><span class="strut" style="height: 0.58333em;"></span><span class="strut bottom" style="height: 0.66666em; vertical-align: -0.08333em;"></span><span class="base"><span class="mord">±</span></span></span>
</span>
60)
</td>
<td>✔️ / ✔️</td>
<td>6h58m20s</td>
</tr>
<tr>
<td><a href="/datasets/musdb.html" class="">MUSDB18</a></td>
<td>2017</td>
<td>(Rafii et al)</td>
<td>150</td>
<td>
(236
<span class="katex">
<span class="katex-mathml">
<math>
<semantics>
<mrow>
<mo>±</mo>
</mrow>
<annotation encoding="application/x-tex">\pm</annotation>
</semantics>
</math>
</span>
<span aria-hidden="true" class="katex-html"><span class="strut" style="height: 0.58333em;"></span><span class="strut bottom" style="height: 0.66666em; vertical-align: -0.08333em;"></span><span class="base"><span class="mord">±</span></span></span>
</span>
95)
</td>
<td>✔️ / ✔️</td>
<td>9h50m</td>
</tr>
</tbody>
</table>
</section>
<section>
<h1>The MUSDB18 dataset</h1>
<ul>
<li>100 train / 50 test full tracks</li>
<li>Mastered with pro. digital audio workstations</li>
<li>compressed STEMS (<code>MUSDB18</code>) and uncompressed WAV <code>MUSDB18-HQ</code>
<li>Parser and Evaluation tools in <i class="fab fa-python"></i></li>
<li><a ref=https://sigsep.github.io/datasets/musdb.html>https://sigsep.github.io/datasets/musdb.html</a></li>
</ul>
<img width="320px" style="float: right" src="assets/ni_logo.png" alt="">
<img width="60%" src="assets/hero.svg" alt="">
</section>
<section>
<h1>Hands on oracle separation</h1>
<h2>Exploring MUSDB18</h2>
<ul>
<li>Start the notebook session</li>
<li>For one track, display waveforms, play some audio</li>
<li>Display spectrogram of mixture</li>
</ul>
<h2>Oracle separation</h2>
<ul>
<li>Get spectrograms of the sources</li>
<li>Display the corresponding soft-mask for vocals</li>
<li>Apply it on the mixture, reconstruct and listen to the result</li>
</ul>
<h2>Performance evaluation</h2>
<ul>
<li>Loop over some musdb tracks</li>
<li>Evaluate oracle separation system on musdb</li>
<li>Compare to state of the art (SiSEC18)</li>
</ul>
</section>
<section>
<h1>A brief history of separation</h1>
<h2>The big picture</h2>
<div class="centered">
<img width="90%" class="stretch" src="assets/discriminative_big_picture.svg" alt="">
</div>
<h6 style="margin-top:3%; color:gray">
Rafii, Zafar, et al. "An Overview of Lead and Accompaniment Separation in Music."
IEEE/ACM TASLP 26.8 (2018): 1307-1335.
</h6>
</section>
<section>
<section>
<h1>A brief history: model-driven methods</h1>
<h2>Harmonicity for the lead</h2>
<img style="float:right" width="70%" src="assets/model_based/ReviewPaper_ Figure2.svg" alt="">
<ul>
<li>Pitch detection</li>
<li>Clean voices</li>
<li>"Metallic" artifacts</li>
</ul>
<!-- an example -->
</section>
<section>
<h1>A brief history: model-driven methods</h1>
<h2>Redundancy for the accompaniment: NMF</h2>
<img style="float:right" width="38%" src="assets/model_based/ReviewPaper_ Figure3.svg" alt="">
<ul>
<li>Spectral templates</li>
<li>Low-rank assumptions</li>
<li>Bad generalization</li>
</ul>
<!-- an example -->
</section>
<section>
<h1>A brief history: model-driven methods</h1>
<h2>Redundancy for the accompaniment: RPCA</h2>
<p>
<img style="margin-top:4%;" width="90%" src="assets/model_based/ReviewPaper_ Figure4.svg" alt=""><p>
<ul>
<li>Low-rank for music</li>
<li>Vocals as unstructured</li>
<li>Strong interferences in general</li>
</ul>
<!-- an example -->
</section>
<section>
<h1>A brief history: model-driven methods</h1>
<h2>Redundancy for the accompaniment: REPET</h2>
<img style="float:right; margin-top:4%" width="65%" src="assets/model_based/ReviewPaper_ Figure5.svg" alt="">
<ul>
<li>Repetitive music</li>
<li>Non-repetitive vocals</li>
<li>Solos in vocals</li>
</ul>
</section>
<section>
<h1>A brief history: model-driven methods</h1>
<h2> Modeling both lead and accompaniment: source filter</h2>
<img style="float:right" width="70%" src="assets/model_based/ReviewPaper_ Figure7.svg" alt="">
<ul>
<li>Harmonic vocals</li>
<li>Low-rank music</li>
<li>Poor generalization</li>
</ul>
</section>
<section>
<h1>A brief history: model-driven methods</h1>
<img style="float:right" width="43%" src="assets/model_based/ReviewPaper_ Figure8.svg" alt="">
<h2>Cascaded methods</h2>
<ul>
<li>Combining methods</li>
<li>Handcrafted systems</li>
<li>Poor generalization</li>
</ul>
</section>
<section>
<h1>A brief history: model-driven methods</h1>
<h2>Fusion of methods</h2>
<img style="float:right" width="49%" src="assets/model_based/ReviewPaper_ Figure9.svg" alt="">
<ul>
<li>Combining in a data-driven way</li>
<li>Doing best than all</li>
<li>Computationally demanding</li>
</ul>
</section>
</section>
<!-- DEEP NEURAL NETWORKS -->
<section>
<h1>A starter on deep neural networks</h1>
<h6 style="margin-top:50%; color:gray">
Y. LeCun, et al. "Deep learning". nature, 521(7553), 436 (2015).
</h6>
</section>
<section>
<h1>Static data</h1>
<h2>The basic fully connected layer</h2>
<img width="80%" src="assets/one_layer.svg" alt="">
<img class="fragment" style="float:right" width="60%" src="assets/non_linearities.svg" alt="">
</section>
<section>
<h1>Static data</h1>
<h2>Basic fully connected network</h2>
<img style="margin-top:5%" width="100%" src="assets/several_layers.svg" alt="">
</section>
<section>
<h1>Static data</h1>
<h2>A usual deep network</h2>
<img style="margin-top:5%" width="100%" src="assets/several_layers_general.svg" alt="">
<ul class="fragment">
<li>Cascading linear and non-linear operations augments expressive power</li>
<li>7 millions parameters in our case</li>
</ul>
</section>
<section>
<h1>Temporal data</h1>
<h6 style="margin-top:50%; color:gray">
colah's blog, <a href="http://colah.github.io/posts/2015-08-Understanding-LSTMs/">Understanding LSTM Networks</a>, 2015.
</h6>
</section>
<section>
<section>
<h1>Temporal data</h1>
<h2>From fully connected to the simple recurrent net</h2>
<img style="float:center" width="90%" src="assets/recurrent_layer_1.svg" alt="">
</section>
<section>
<h1>Temporal data</h1>
<h2>From fully connected to the simple recurrent net</h2>
<img style="float:center" width="90%" src="assets/recurrent_layer_2.svg" alt="">
</section>
<section>
<h1>Temporal data</h1>
<h2>From fully connected to the simple recurrent net</h2>
<img style="float:center" width="90%" src="assets/time_distributed_dense.svg" alt="">
</section>
<section>
<h1>Temporal data</h1>
<h2>From fully connected to the simple recurrent net</h2>
<img style="float:center" width="90%" src="assets/recurrent_layer_3.svg" alt="">
</section>
<section>
<h1>Temporal data</h1>
<h2>The simple recurrent net</h2>
<img style="float:left" width="24%" src="assets/recurrent_layer_4.svg" alt="">
<ul>
<li>$y_{t}=f\left(linear\left\{ x_{t},y_{t-1}\right\} \right)$</li>
<li>Similar to a Markov model
<ul>
<li>Exponential decay of information</li>
<li>Vanishing or exploding gradient for training</li>
</ul></li>
</ul>
<p>
<ul>
<li>Limited for long-term dependencies</li>
</ul>
<h6 style="margin-top:30%; color:gray">
P. Huang, et al. "Deep learning for monaural speech separation". (2014) ICASSP.
</h6>
</section>
</section>
<section>
<section>
<h1>Temporal data</h1>
<h2>The long short term memory (LSTM)</h2>
<img style="margin-top:2%; float:center" width="100%" src="assets/lstm_1.svg" alt="">
</section>
<section>
<h1>Temporal data</h1>
<h2>The long short term memory (LSTM)</h2>
<img style="margin-top:2%; float:center" width="100%" src="assets/lstm_2.svg" alt="">
</section>
<section>
<h1>Temporal data</h1>
<h2>The long short term memory (LSTM)</h2>
<img style="margin-top:2%; float:center" width="80%" src="assets/lstm_3.svg" alt="">
</section>
</section>
<section>
<section>
<h1>Temporal data</h1>
<h2>The bi-LSTM</h2>
<img style="margin-top:2%; float:left" width="50%" src="assets/blstm_1.svg" alt="">
<ul>
<li>LSTM are causal systems</li>
<li>Predicts future from past</li>
</ul>
</section>
<section>
<h1>Temporal data</h1>
<h2>The bi-LSTM</h2>
<img style="margin-top:2%; float:left" width="50%" src="assets/blstm_2.svg" alt="">
<ul>
<li>We can use anti-causal LSTM</li>
<li>Different predictions!</li>
</ul>
</section>
<section>
<h1>Temporal data</h1>
<h2>The bi-LSTM</h2>
<img style="margin-top:2%; float:left" width="50%" src="assets/blstm_3.svg" alt="">
<ul>
<li>Independent forward and backward</li>
<li>Outputs can be concatenated</li>
<li class="fragment">Outputs can be summed<p>
<img style="margin-top:2%; float:center" width="100%" src="assets/combine_blstm.svg" alt="">
</li>
</ul>
</section>
</section>
<section>
<h1>Discriminative source separation: one source</h1>
<img width="48%" style="margin-top:5%; float:left" src="assets/discriminative_main_lines.svg" alt="">
<img width="20%" style="margin-top:7%; float:left" src="assets/vanilla.svg" alt="">
<ul style="margin-top:3%;">
<li>Directly get source from mixture</li>
<li>Straightforward inference</li>
<li>Trained on paired mixtures/sources</li>
</ul>
</section>
<section>
<h1>Discriminative source separation: all sources</h1>
<img width="100%" style="float:center" src="assets/vanilla_lstm_test.svg" alt="">
<h6 style="margin-top:4%; color:gray">
S. Uhlich, et al. "Improving music source separation based on deep neural networks through data augmentation and network blending." (2017) ICASSP.
</h6>
</section>
<section>
<h1>Generative source separation</h1>
<img width="48%" style="margin-top:5%; float:left" src="assets/generative_model.svg" alt="">
<ul style="margin-top:5%;">
<li>The model can transform random noise to realistic spectrograms</li>
<li>Training is done on sources only, without mixtures</li>
</ul>
</section>
<section>
<h1>Inference for generative separation</h1>
<img width="100%" style="margin-top:1%; float:left" src="assets/generative_test.svg" alt="">
<ul style="margin-top:5%;">
<li>Testing requires inference of the latent variables ("noise")</li>
</ul>
</section>
<section>
<style>
.container{display: flex;}
.col1{flex: 3;}
.col2{flex: 1;}
</style>
<h1>Inference for generative separation: recent pointers</h1>
<div class="container">
<div class="col1">
<ul>Rejection sampling with Variational autoencoders
<li style="color:gray">
S. Leglaive et al. "A variance modeling framework based on variational autoencoders for speech enhancement" MLSP, 2018.
</li>
<li style="color:gray">
Y. Bando et al. "Statistical speech enhancement based on probabilistic integration of variational autoencoder and non-negative matrix factorization", ICASSP 2018
</li>
</ul>
<p>
<ul>Rejection sampling with GANs
<li style="color:gray">
Y. Subakan et al. "Generative adversarial source separation", ICASSP 2018.
</li>
</ul>
<p>
<ul>Inference with encoder networks
<li style="color:gray">
M. Pariente et al. "A statistically principled and computationally efficient approach to speech enhancement using variational autoencoders" arXiv, 2019.
</li>
</ul>
</div>
<div class="col2">
<img width="100%" style="margin-top:1%; float:right" src="assets/iterative_generative_test.svg" alt="">
</div>
</div>
</section>
<section>
<h1>Hands on using pre-trained <code>open-unmix</code></h1>
<ul style="margin-top:10%">
<li>Load the pre-trained <code>open-unmix</code></li>
<p>
<li>Separate a MUSDB7 track</li>
<p>
<li>Compute scores and compare with oracle</li>
</ul>
<p>
</section>
<section>
<h1>Training a DNN</h1>
<ul>
<li>Vocabulary</li>
<li>Gradient descent</li>
<li>Discriminative training</li>
<li>Generative training</li>
</ul>
</section>
<section>
<h1>Vocabulary</h1>
<img style="margin-top:5%" width="100%" src="assets/training_dataset.svg" alt="">
</section>
<section>
<h1>Vocabulary</h1>
<img style="margin-top:5%" width="100%" src="assets/training_dataset_batch.svg" alt="">
</section>
<section>
<h1>Vocabulary</h1>
<img style="margin-top:5%" width="100%" src="assets/training_dataset_learning.svg" alt="">
</section>
<section>
<h1>Vocabulary</h1>
<img style="margin-top:5%" width="100%" src="assets/one_epoch.gif" alt="">
</section>
<section>
<h1>Gradient descent</h1>
<img style="margin-top:5%; float:right" width="43%" src="assets/gradient_descent.gif" alt="">
<ul>
<li>$loss\leftarrow \sum_{(x,y)\in batch}cost\left(y_\Theta\left(x\right), y\right)$</li>
<li>Update $\Theta$ to reduce the loss!</li>
<p>
<li>We can compute $\frac{\partial loss}{\partial\Theta_{i}}$ for any parameter $\Theta_i$
<ul>
<li>"The influence of $\Theta_i$ on the error"</li>
<li>It's the <strong>gradient</strong></li>
<li>Computed through <strong>backpropagation</strong></li>
</ul>
</li><p>
<li>A simple optimization: $\Theta_i\leftarrow \Theta_i - \lambda \frac{\partial loss}{\partial\Theta_{i}}$
<ul>
<li>It's the <strong>stochastic gradient descent</strong></li>
<li>$\lambda$ is the <strong>learning rate</strong></li>
</ul>
</li>
<li>Batching is important</li>
</ul>
<p>
There are many other optimization algorithms...
<!--
<h6 style="color:gray">
<ul>
<li><strong>RMSprop</strong> T. Tieleman, et al. "Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude." COURSERA: Neural networks for machine learning 4.2 (2012): 26-31.
<li><strong>Adam</strong> D. Kingma, et al. "Adam: A method for stochastic optimization." arXiv preprint (2014) arXiv:1412.6980.
</ul>
</h6> -->
</section>
<section>
<h1>Gradient descent</h1>
<h2>Learning rate wisdom</h2>
<img style="margin-top:5%; " width="45%" src="assets/gradient_descent_big.gif" alt="">
<!-- <img style="margin-top:5%; float:center" width="33%" src="assets/gradient_descent.gif" alt=""> -->
<img style="margin-top:5%; " width="40%" src="assets/learningrates.jpg" alt="">
<h6 style="margin-top:10%; color:gray">
Leonardo Araujo dos Santos, <a href="https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/model_optimization.html">Artificial Intelligence</a>, 2017.
</h6>
</section>
<section>
<h1>Gradient descent</h1>
<h2>Regularization with dropout</h2>
<img style="margin-top:5%; float:left" width="30%" src="assets/dropout.gif" alt="">
<ul style="margin-top:5%;">
<li>Parts of the net randomly set to 0</li>
<li>No unit should be critical: <i>regularization</i></li>
<li>Probabilistic interpretation</li>
</ul>
<h6 style="margin-top:28%; color:gray">
N Srivastava, et al. "Dropout: a simple way to prevent neural networks from overfitting". JMLR. (2014) 15(1), 1929-1958.
</h6>
</section>
<section>
<h1>Gradient descent</h1>
<h2>Data augmentation</h2>
<ul>
<li>Artificially increase the size of the dataset</li>
<p>
<li>An active research topic in audio
<ul style="color:gray">
<li>S. Uhlich, et al. "Improving music source separation based on deep neural networks through data augmentation and network blending." (2017) ICASSP</li>
<li>A. Cohen-Hadria, et al. "Improving singing voice separation using Deep U-Net and Wave-U-Net with data augmentation."" arXiv 2019.
</ul>
</li>