File size: 40,015 Bytes
b386992
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4",
      "toc_visible": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU",
    "gpuClass": "standard"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "# Introduction"
      ],
      "metadata": {
        "id": "rtBDkKqVGZJ8"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "In this tutorial, we will prepare a dataset using our [TTS Dataset Processing Scripts](https://github.com/NVIDIA/NeMo/tree/main/scripts/dataset_processing/tts) and use it for training a FastPitch model.\n",
        "\n",
        "**This tutorial uses a different workflow than all other existing TTS tutorials. The scripts and classes used are all experimental and not yet ready for production**."
      ],
      "metadata": {
        "id": "pZ2QSsXuGbMe"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "# License"
      ],
      "metadata": {
        "id": "7X-TwhdTGmlc"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "> Copyright (c) 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.\n",
        ">\n",
        "> Licensed under the Apache License, Version 2.0 (the \"License\"); you may not use this file except in compliance with the License. You may obtain a copy of the License at\n",
        ">\n",
        "> http://www.apache.org/licenses/LICENSE-2.0\n",
        ">\n",
        "> Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an \"AS IS\" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License."
      ],
      "metadata": {
        "id": "fCQUeZRPGnoe"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Install"
      ],
      "metadata": {
        "id": "3OZassNG5xff"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "BRANCH = 'main'\n",
        "NEMO_ROOT_DIR = '/content/nemo'"
      ],
      "metadata": {
        "id": "QLLoj7bD0W5f"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "WZvQvPkIhRi3"
      },
      "outputs": [],
      "source": [
        "# Install NeMo library. If you are running locally (rather than on Google Colab), comment out the below lines\n",
        "# and instead follow the instructions at https://github.com/NVIDIA/NeMo#Installation\n",
        "!python -m pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[all]"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "\n",
        "# Download local version of NeMo scripts. If you are running locally and want to use your own local NeMo code,\n",
        "# comment out the below lines and set NEMO_ROOT_DIR to your local path.\n",
        "!git clone -b $BRANCH https://github.com/NVIDIA/NeMo.git $NEMO_ROOT_DIR"
      ],
      "metadata": {
        "id": "tvsgWO_WhV3M"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Dataset Preparation"
      ],
      "metadata": {
        "id": "fM4QPsLTnzK7"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "For our tutorial, we use a subset of [VCTK](https://datashare.ed.ac.uk/handle/10283/2950) dataset with 5 speakers (p225-p229)."
      ],
      "metadata": {
        "id": "tkZC6Dl7KRl6"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import os\n",
        "import tarfile\n",
        "import wget\n",
        "from pathlib import Path\n",
        "\n",
        "from nemo.collections.asr.parts.utils.manifest_utils import read_manifest, write_manifest"
      ],
      "metadata": {
        "id": "sYzvAYr2vo1K"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Configure nemo paths\n",
        "NEMO_DIR = Path(NEMO_ROOT_DIR)\n",
        "NEMO_EXAMPLES_DIR = NEMO_DIR / \"examples\" / \"tts\"\n",
        "NEMO_CONFIG_DIR = NEMO_EXAMPLES_DIR / \"conf\"\n",
        "NEMO_SCRIPT_DIR = NEMO_DIR / \"scripts\" / \"dataset_processing\" / \"tts\""
      ],
      "metadata": {
        "id": "APo1m5M-v3pB"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Create dataset directory\n",
        "root_dir = Path(\"/content\")\n",
        "data_root = root_dir / \"data\"\n",
        "\n",
        "data_root.mkdir(parents=True, exist_ok=True)"
      ],
      "metadata": {
        "id": "aoxN1QsUzX-k"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Download the dataset\n",
        "dataset_url = \"https://vctk-subset.s3.amazonaws.com/vctk_subset_multispeaker.tar.gz\"\n",
        "dataset_tar_filepath = data_root / \"vctk.tar.gz\"\n",
        "\n",
        "if not os.path.exists(dataset_tar_filepath):\n",
        "    wget.download(dataset_url, out=str(dataset_tar_filepath))"
      ],
      "metadata": {
        "id": "mArlQd5Hk36b"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Extract the dataset\n",
        "with tarfile.open(dataset_tar_filepath) as tar_f:\n",
        "    tar_f.extractall(data_root)"
      ],
      "metadata": {
        "id": "p987cjtOy9C7"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "DATA_DIR = data_root / \"vctk_subset_multispeaker\""
      ],
      "metadata": {
        "id": "Ko6dxYJW0i3G"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Visualize the raw dataset\n",
        "train_raw_filepath = DATA_DIR / \"train.json\"\n",
        "!head $train_raw_filepath"
      ],
      "metadata": {
        "id": "We5FHYQt5BeO"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Manifest Processing"
      ],
      "metadata": {
        "id": "i3jsk2HCMSU5"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "The downloaded manifest uses our traditional format for TTS training. The scripts here require it to be formatted slightly differently.\n",
        "\n",
        "The `speaker` field used to be an *integer* ID corresponding to an array index that the FastPitch model would query. Now we represent it as a *string* so we can give each speaker a human-friendly name. The mapping from speaker name to speaker index will be provided at training time.\n",
        "\n",
        "As a best practice, we suggest prepending the `speaker` field with the name of the dataset so that it is guaranteed to be unique across all datasets (eg. *vctk_225*, instead of *225*).\n",
        "\n",
        "The `audio_filepath` field used to require an *absolute path* which had to be manually updated depending on where the dataset was on your computer. Absolute paths still work, but now you can optionally provide it as a *relative path*, with the root directory provided as an argument to each script."
      ],
      "metadata": {
        "id": "N8WuAGJsMHRn"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "def update_metadata(data_type):\n",
        "    input_filepath = DATA_DIR / f\"{data_type}.json\"\n",
        "    output_filepath = DATA_DIR / f\"{data_type}_raw.json\"\n",
        "\n",
        "    entries = read_manifest(input_filepath)\n",
        "    for entry in entries:\n",
        "        # Provide relative path instead of absolute path\n",
        "        entry[\"audio_filepath\"] = entry[\"audio_filepath\"].replace(\"audio/\", \"\")\n",
        "        # Prepend speaker ID with the name of the dataset: 'vctk'\n",
        "        entry[\"speaker\"] = f\"vctk_{entry['speaker']}\"\n",
        "\n",
        "    write_manifest(output_path=output_filepath, target_manifest=entries, ensure_ascii=False)"
      ],
      "metadata": {
        "id": "zoCRrKQ20VZP"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "update_metadata(\"dev\")\n",
        "update_metadata(\"train\")"
      ],
      "metadata": {
        "id": "PaCc3GCG1UbH"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Visualize updated 'audio_filepath' and 'speaker' fields\n",
        "train_filepath = DATA_DIR / \"train_raw.json\"\n",
        "!head $train_filepath"
      ],
      "metadata": {
        "id": "bVLIB3Ip1Aqn"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Text Preprocessing"
      ],
      "metadata": {
        "id": "e3jHTOhL1M5_"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "First we will process the text transcripts using the script [preprocess_text.py](https://github.com/NVIDIA/NeMo/blob/main/scripts/dataset_processing/tts/preprocess_text.py).\n",
        "\n",
        "This step mainly passes the text through our NeMo *text normalizer* and then stores the output in the `normalized_text` field. It also has a few optional transformations, such as lowercasing the text."
      ],
      "metadata": {
        "id": "H2rYykFLSR5t"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "text_preprocessing_script = NEMO_SCRIPT_DIR / \"preprocess_text.py\"\n",
        "\n",
        "# Number of threads to parallelize text processing across\n",
        "num_workers = 4\n",
        "# Text normalizer to apply\n",
        "normalizer_config_filepath = NEMO_CONFIG_DIR / \"text\" / \"normalizer_en.yaml\"\n",
        "# Whether to lowercase output text. We can safely do this here because we will train on IPA phonemes.\n",
        "# If training on graphemes only, then consider disabling this to leave text with its original capitalization.\n",
        "lower_case = True\n",
        "# Whether to overwrite output manifest, if it exists\n",
        "overwrite_manifest = True\n",
        "# Batch size for joblib parallelization. Increasing this value might speed up the script, depending on your CPU.\n",
        "joblib_batch_size = 16\n",
        "\n",
        "# Python wrapper to invoke the given bash script with the given input args\n",
        "def run_script(script, args):\n",
        "    args = ' \\\\'.join(args)\n",
        "    cmd = f\"python {script} \\\\{args}\"\n",
        "\n",
        "    print(cmd.replace(\" \\\\\", \"\\n\"))\n",
        "    print()\n",
        "    !$cmd\n",
        "\n",
        "def preprocess_text(data_type):\n",
        "    input_filepath = DATA_DIR / f\"{data_type}_raw.json\"\n",
        "    output_filepath = DATA_DIR / f\"{data_type}_text.json\"\n",
        "\n",
        "    args = [\n",
        "        f\"--input_manifest={input_filepath}\",\n",
        "        f\"--output_manifest={output_filepath}\",\n",
        "        f\"--num_workers={num_workers}\",\n",
        "        f\"--normalizer_config_path={normalizer_config_filepath}\",\n",
        "        f\"--joblib_batch_size={joblib_batch_size}\"\n",
        "    ]\n",
        "    if lower_case:\n",
        "      args.append(\"--lower_case\")\n",
        "    if overwrite_manifest:\n",
        "        args.append(\"--overwrite\")\n",
        "\n",
        "    run_script(text_preprocessing_script, args)"
      ],
      "metadata": {
        "id": "6Z1vRsPd0g2s"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "preprocess_text(\"dev\")"
      ],
      "metadata": {
        "id": "qg6iK3NyrZvx"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "preprocess_text(\"train\")"
      ],
      "metadata": {
        "id": "DkLhSL_n1QAS"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Visualize the output of the 'normalized_text' field.\n",
        "train_text_filepath = DATA_DIR / \"train_text.json\"\n",
        "!head $train_text_filepath"
      ],
      "metadata": {
        "id": "6qHbl0Cf5kQn"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Audio Preprocessing"
      ],
      "metadata": {
        "id": "alrRDWio41qi"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Next we process the audio data using [preprocess_audio.py](https://github.com/NVIDIA/NeMo/blob/main/scripts/dataset_processing/tts/preprocess_audio.py).\n",
        "\n",
        "During this step we apply the following transformations:\n",
        "\n",
        "1. Resample the audio from 48khz to 44.1khz so that it is compatible with our default training configuration.\n",
        "2. Remove long silence from the beginning and end of each audio file. This can be done using an *energy* based approach which will work on clean audio, or using *voice activity detection (VAD)* which also works on audio with background or static noise (eg. from a microphone).\n",
        "3. Scale the audio so that files have approximately the same volume level.\n",
        "4. Filter out audio files which are too long or too short.\n",
        "\n"
      ],
      "metadata": {
        "id": "4WfEaMwpUsFt"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import IPython.display as ipd"
      ],
      "metadata": {
        "id": "WEvIefjnd7AG"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "audio_preprocessing_script = NEMO_SCRIPT_DIR / \"preprocess_audio.py\"\n",
        "\n",
        "# Directory with raw audio data\n",
        "input_audio_dir = DATA_DIR / \"audio\"\n",
        "# Directory to write preprocessed audio to\n",
        "output_audio_dir = DATA_DIR / \"audio_preprocessed\"\n",
        "# Whether to overwrite existing audio, if it exists in the output directory\n",
        "overwrite_audio = True\n",
        "# Whether to overwrite output manifest, if it exists\n",
        "overwrite_manifest = True\n",
        "# Number of threads to parallelize audio processing across\n",
        "num_workers = 4\n",
        "# Downsample data from 48khz to 44.1khz for compatibility\n",
        "output_sample_rate = 44100\n",
        "# Format of output audio files. Use \"flac\" to compress to a smaller file size.\n",
        "output_format = \"flac\"\n",
        "# Method for silence trimming. Can use \"energy.yaml\" or \"vad.yaml\".\n",
        "# We use VAD for VCTK because the audio has background noise.\n",
        "trim_config_path = NEMO_CONFIG_DIR / \"trim\" / \"vad.yaml\"\n",
        "# Volume level (0, 1] to normalize audio to\n",
        "volume_level = 0.95\n",
        "# Filter out audio shorter than min_duration or longer than max_duration seconds.\n",
        "# We set these bounds relatively low/high, as we can place stricter limits at training time\n",
        "min_duration = 0.25\n",
        "max_duration = 30.0\n",
        "# Output file with entries that are filtered out based on duration\n",
        "filter_file = DATA_DIR / \"filtered.json\"\n",
        "\n",
        "def preprocess_audio(data_type):\n",
        "    input_filepath = DATA_DIR / f\"{data_type}_text.json\"\n",
        "    output_filepath = DATA_DIR / f\"{data_type}_manifest.json\"\n",
        "\n",
        "    args = [\n",
        "        f\"--input_manifest={input_filepath}\",\n",
        "        f\"--output_manifest={output_filepath}\",\n",
        "        f\"--input_audio_dir={input_audio_dir}\",\n",
        "        f\"--output_audio_dir={output_audio_dir}\",\n",
        "        f\"--num_workers={num_workers}\",\n",
        "        f\"--output_sample_rate={output_sample_rate}\",\n",
        "        f\"--output_format={output_format}\",\n",
        "        f\"--trim_config_path={trim_config_path}\",\n",
        "        f\"--volume_level={volume_level}\",\n",
        "        f\"--min_duration={min_duration}\",\n",
        "        f\"--max_duration={max_duration}\",\n",
        "        f\"--filter_file={filter_file}\",\n",
        "    ]\n",
        "    if overwrite_manifest:\n",
        "        args.append(\"--overwrite_manifest\")\n",
        "    if overwrite_audio:\n",
        "        args.append(\"--overwrite_audio\")\n",
        "\n",
        "    run_script(audio_preprocessing_script, args)"
      ],
      "metadata": {
        "id": "0kQ1UDnGfdX6"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "preprocess_audio(\"dev\")"
      ],
      "metadata": {
        "id": "ai0zbXSOriuY"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "preprocess_audio(\"train\")"
      ],
      "metadata": {
        "id": "NUKnidQYfgDo"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "We should listen to a few audio files before and after the processing so be sure we configured it correctly.\n",
        "\n",
        "Note that the processed audio is louder. It is also shorter because we trimmed the leading and trailing silence."
      ],
      "metadata": {
        "id": "x2yhJtsj2lDR"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "audio_file = \"p228_009.wav\"\n",
        "audio_filepath = input_audio_dir / audio_file\n",
        "processed_audio_filepath = output_audio_dir / audio_file.replace(\".wav\", \".flac\")\n",
        "\n",
        "print(\"Original audio.\")\n",
        "ipd.display(ipd.Audio(audio_filepath))\n",
        "\n",
        "print(\"Processed audio.\")\n",
        "ipd.display(ipd.Audio(processed_audio_filepath))"
      ],
      "metadata": {
        "id": "_fM3GwJxkjOA"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Speaker Mapping"
      ],
      "metadata": {
        "id": "d129p0nrr3PD"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "We can use [create_speaker_map.py](https://github.com/NVIDIA/NeMo/blob/main/scripts/dataset_processing/tts/create_speaker_map.py) to easily create a mapping from speaker ID strings to integer indices that will be used at training time.\n",
        "\n",
        "The script will simply sort the speaker IDs and assign them numbers `[0, num_speakers)` in alphabetical order."
      ],
      "metadata": {
        "id": "ZJ1MWX3F3X9u"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "speaker_map_script = NEMO_SCRIPT_DIR / \"create_speaker_map.py\"\n",
        "\n",
        "train_manifest_filepath = DATA_DIR / \"train_manifest.json\"\n",
        "dev_manifest_filepath = DATA_DIR / \"dev_manifest.json\"\n",
        "speaker_filepath = DATA_DIR / \"speakers.json\"\n",
        "\n",
        "args = [\n",
        "    f\"--manifest_path={train_manifest_filepath}\",\n",
        "    f\"--manifest_path={dev_manifest_filepath}\",\n",
        "    f\"--speaker_map_path={speaker_filepath}\"\n",
        "]\n",
        "\n",
        "run_script(speaker_map_script, args)"
      ],
      "metadata": {
        "id": "b5gdccYhr5Gk"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Visualize the speaker map file.\n",
        "!head $speaker_filepath"
      ],
      "metadata": {
        "id": "CMcC2Nqmt5AR"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Feature Computation"
      ],
      "metadata": {
        "id": "jyFxOjy6t8vo"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Before training FastPitch, we need to compute some features for every audio file. The default [config file](https://github.com/NVIDIA/NeMo/blob/main/examples/tts/conf/feature/feature_44100.yaml) we will use has parameters for computing the **pitch** and **energy** of every audio frame. Be default it will also compute a **voiced_mask** indicating which audio frames have no pitch (eg. because they contain silence)."
      ],
      "metadata": {
        "id": "QNPpwkM49orB"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "feature_script = NEMO_SCRIPT_DIR / \"compute_features.py\"\n",
        "\n",
        "sample_rate = 44100\n",
        "\n",
        "if sample_rate == 22050:\n",
        "    feature_config_filename = \"feature_22050.yaml\"\n",
        "elif sample_rate == 44100:\n",
        "    feature_config_filename = \"feature_44100.yaml\"\n",
        "else:\n",
        "    raise ValueError(f\"Unsupported sampling rate {sample_rate}\")\n",
        "\n",
        "feature_config_path = NEMO_CONFIG_DIR / \"feature\" / feature_config_filename\n",
        "audio_dir = DATA_DIR / \"audio_preprocessed\"\n",
        "feature_dir = DATA_DIR / \"features\"\n",
        "num_workers = 4\n",
        "\n",
        "def compute_features(data_type):\n",
        "    input_filepath = DATA_DIR / f\"{data_type}_manifest.json\"\n",
        "\n",
        "    args = [\n",
        "        f\"--feature_config_path={feature_config_path}\",\n",
        "        f\"--manifest_path={input_filepath}\",\n",
        "        f\"--audio_dir={audio_dir}\",\n",
        "        f\"--feature_dir={feature_dir}\",\n",
        "        f\"--num_workers={num_workers}\"\n",
        "    ]\n",
        "\n",
        "    run_script(feature_script, args)"
      ],
      "metadata": {
        "id": "AI4aLRFbt_NQ"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "compute_features(\"dev\")"
      ],
      "metadata": {
        "id": "kQqPw3uRwEsO"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "compute_features(\"train\")"
      ],
      "metadata": {
        "id": "ct1fN_4pwCu9"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "The features are stored in the specified `feature_dir`."
      ],
      "metadata": {
        "id": "db83_UcOCOIo"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!ls $feature_dir"
      ],
      "metadata": {
        "id": "_8bHP4j56LWG"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Feature Statistics"
      ],
      "metadata": {
        "id": "QsuxK1P0x7hZ"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "For training it is beneficial for us to *normalize* our features. The most standard approach is to apply *mean-variance normalization* so that each feature has a mean of 0 and variance of 1. To do this we need to compute the *dataset statistics* with the mean and variance of each feature.\n",
        "\n",
        "For TTS it also helps\n",
        "*   Normalize features using speaker-level statistics.\n",
        "*   Use the `voiced_mask` to set the feature values of non-voiced audio frames to 0.\n",
        "\n",
        "Using the [compute_feature_stats.py](https://github.com/NVIDIA/NeMo/blob/main/scripts/dataset_processing/tts/compute_feature_stats.py) script we will compute the mean and variance of each feature for each speaker. The input to the script is the same [config file](https://github.com/NVIDIA/NeMo/blob/main/examples/tts/conf/feature/feature_44100.yaml) we used to compute the features."
      ],
      "metadata": {
        "id": "O8GiAnAMCNeh"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "feature_stats_script = NEMO_SCRIPT_DIR / \"compute_feature_stats.py\"\n",
        "\n",
        "train_manifest_filepath = DATA_DIR / \"train_manifest.json\"\n",
        "output_stats_path = DATA_DIR / \"feature_stats.json\"\n",
        "\n",
        "args = [\n",
        "    f\"--feature_config_path={feature_config_path}\",\n",
        "    f\"--manifest_path={train_manifest_filepath}\",\n",
        "    f\"--audio_dir={audio_dir}\",\n",
        "    f\"--feature_dir={feature_dir}\",\n",
        "    f\"--stats_path={output_stats_path}\"\n",
        "]\n",
        "\n",
        "run_script(feature_stats_script, args)"
      ],
      "metadata": {
        "id": "DC4c1L3CxH-h"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "The output feature statistics file contains the mean and variance of the pitch and energy for the entire dataset (under the key `global`), and for each speaker in the dataset."
      ],
      "metadata": {
        "id": "zos96yaoFho1"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!head $output_stats_path"
      ],
      "metadata": {
        "id": "fOz1cpIdFcG9"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# HiFi-GAN Training"
      ],
      "metadata": {
        "id": "oRO842MUyODC"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Our standard FastPitch model is a two-part recipe consisting of the **FastPitch** acoustic model which predicts a mel spectrogram from text, and **HiFi-GAN** vocoder which predicts audio from the mel spectrogram.\n",
        "\n",
        "We will train HiFi-GAN first so that we can use it to help evaluate the performance of FastPitch as it is being trained.\n",
        "\n",
        "HiFi-GAN training only requires a manifest with the `audio_filepath` field. All other fields in the manifest are for FastPitch training.\n",
        "\n",
        "Here we show how to train these models from scratch. You can also fine-tune them from pretrained checkpoints as mentioned in our [FastPitch fine-tuning tutorial](https://github.com/NVIDIA/NeMo/blob/main/tutorials/tts/FastPitch_Finetuning.ipynb), but pretrained checkpoints compatible with these experimental recipes are not yet available on NGC.\n"
      ],
      "metadata": {
        "id": "E4wUKYOfH8ax"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import torch"
      ],
      "metadata": {
        "id": "pqfl9jAYMJob"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "dataset_name = \"vctk\"\n",
        "audio_dir = DATA_DIR / \"audio_preprocessed\"\n",
        "train_manifest_filepath = DATA_DIR / \"train_manifest.json\"\n",
        "dev_manifest_filepath = DATA_DIR / \"dev_manifest.json\""
      ],
      "metadata": {
        "id": "jK2rr-Kr6Qg8"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "hifigan_training_script = NEMO_EXAMPLES_DIR / \"hifigan.py\"\n",
        "\n",
        "# The total number of training steps will be (epochs * steps_per_epoch)\n",
        "epochs = 10\n",
        "steps_per_epoch = 10\n",
        "\n",
        "sample_rate = 44100\n",
        "\n",
        "# Config files specifying all HiFi-GAN parameters\n",
        "hifigan_config_dir = NEMO_CONFIG_DIR / \"hifigan_dataset\"\n",
        "\n",
        "if sample_rate == 22050:\n",
        "    hifigan_config_filename = \"hifigan_22050.yaml\"\n",
        "elif sample_rate == 44100:\n",
        "    hifigan_config_filename = \"hifigan_44100.yaml\"\n",
        "else:\n",
        "    raise ValueError(f\"Unsupported sampling rate {sample_rate}\")\n",
        "\n",
        "# Name of the experiment that will determine where it is saved locally and in TensorBoard and WandB\n",
        "run_id = \"test_run\"\n",
        "exp_dir = root_dir / \"exps\"\n",
        "hifigan_exp_output_dir = exp_dir / \"HifiGan\" / run_id\n",
        "# Directory where predicted audio will be stored periodically throughout training\n",
        "hifigan_log_dir = hifigan_exp_output_dir / \"logs\"\n",
        "\n",
        "if torch.cuda.is_available():\n",
        "    accelerator=\"gpu\"\n",
        "    batch_size = 16\n",
        "else:\n",
        "    accelerator=\"cpu\"\n",
        "    batch_size = 2\n",
        "\n",
        "args = [\n",
        "    f\"--config-path={hifigan_config_dir}\",\n",
        "    f\"--config-name={hifigan_config_filename}\",\n",
        "    f\"max_epochs={epochs}\",\n",
        "    f\"weighted_sampling_steps_per_epoch={steps_per_epoch}\",\n",
        "    f\"batch_size={batch_size}\",\n",
        "    f\"log_dir={hifigan_log_dir}\",\n",
        "    f\"exp_manager.exp_dir={exp_dir}\",\n",
        "    f\"+exp_manager.version={run_id}\",\n",
        "    f\"trainer.accelerator={accelerator}\",\n",
        "    f\"+train_ds_meta.{dataset_name}.manifest_path={train_manifest_filepath}\",\n",
        "    f\"+train_ds_meta.{dataset_name}.audio_dir={audio_dir}\",\n",
        "    f\"+val_ds_meta.{dataset_name}.manifest_path={dev_manifest_filepath}\",\n",
        "    f\"+val_ds_meta.{dataset_name}.audio_dir={audio_dir}\",\n",
        "    f\"+log_ds_meta.{dataset_name}.manifest_path={dev_manifest_filepath}\",\n",
        "    f\"+log_ds_meta.{dataset_name}.audio_dir={audio_dir}\"\n",
        "]"
      ],
      "metadata": {
        "id": "Vr4D-NB-yQx8"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# If an error occurs, log the entire stacktrace.\n",
        "os.environ[\"HYDRA_FULL_ERROR\"] = \"1\""
      ],
      "metadata": {
        "id": "Bn8lQG0PxWGi"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "run_script(hifigan_training_script, args)"
      ],
      "metadata": {
        "id": "yUxFCNrE3Ywi"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "During training, the model will automatically save predictions for all files specified in the `log_ds_meta` manifest."
      ],
      "metadata": {
        "id": "BBPIpS-lL6z9"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "hifigan_log_epoch_dir = hifigan_log_dir / \"epoch_10\" / dataset_name\n",
        "!ls $hifigan_log_epoch_dir"
      ],
      "metadata": {
        "id": "rSFOm1Sg46Lh"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "This makes it easy to listen to the audio to determine how well the model is performing. We can decide to stop training when either:\n",
        "\n",
        "*   The predicted audio sounds almost exactly the same as the original audio\n",
        "*   The predicted audio stops improving in between epochs.\n",
        "\n",
        "**Note that the dataset in this tutorial is too small to get good quality audio output.**"
      ],
      "metadata": {
        "id": "oCJs7oCLMIjD"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "audio_filepath = hifigan_log_epoch_dir / \"p225_143.wav\"\n",
        "ipd.display(ipd.Audio(audio_filepath))"
      ],
      "metadata": {
        "id": "G6k4ymzfJ5Y6"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# FastPitch Training"
      ],
      "metadata": {
        "id": "lV--2Wph7NPG"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Finally we can train the FastPitch model itself. The FastPitch training recipe requires:\n",
        "\n",
        "1. Training manifest(s) with `audio_filepath` and `text` or `normalized_text` fields.\n",
        "2. Precomputed features such as *pitch* and *energy* specified in the feature [config file](https://github.com/NVIDIA/NeMo/blob/main/examples/tts/conf/feature/feature_44100.yaml).\n",
        "3. (Optional) Statistics file for normalizing features.\n",
        "4. (Optional) For a multi-speaker model, the manifest needs a `speaker` field and JSON file mapping speaker IDs to speaker indices.\n",
        "5. (Optional) To train with IPA phonemes, a [phoneme dictionary](https://github.com/NVIDIA/NeMo/blob/main/scripts/tts_dataset_files/ipa_cmudict-0.7b_nv23.01.txt) and optional [heteronyms file](https://github.com/NVIDIA/NeMo/blob/main/scripts/tts_dataset_files/heteronyms-052722)\n",
        "6. (Optional) HiFi-GAN checkpoint or [NGC model name](https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/tts/models/hifigan.py#L413) for generating audio predictions during training.\n",
        "\n"
      ],
      "metadata": {
        "id": "aOuoPXDhOVD7"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "fastpitch_training_script = NEMO_EXAMPLES_DIR / \"fastpitch.py\"\n",
        "\n",
        "# The total number of training steps will be (epochs * steps_per_epoch)\n",
        "epochs = 10\n",
        "steps_per_epoch = 10\n",
        "\n",
        "num_speakers = 5\n",
        "sample_rate = 44100\n",
        "\n",
        "# Config files specifying all FastPitch parameters\n",
        "fastpitch_config_dir = NEMO_CONFIG_DIR / \"fastpitch\"\n",
        "\n",
        "if sample_rate == 22050:\n",
        "    fastpitch_config_filename = \"fastpitch_22050.yaml\"\n",
        "elif sample_rate == 44100:\n",
        "    fastpitch_config_filename = \"fastpitch_44100.yaml\"\n",
        "else:\n",
        "    raise ValueError(f\"Unsupported sampling rate {sample_rate}\")\n",
        "\n",
        "# Metadata files and directories\n",
        "dataset_file_dir = NEMO_DIR / \"scripts\" / \"tts_dataset_files\"\n",
        "phoneme_dict_path = dataset_file_dir / \"ipa_cmudict-0.7b_nv23.01.txt\"\n",
        "heteronyms_path = dataset_file_dir / \"heteronyms-052722\"\n",
        "\n",
        "speaker_path = DATA_DIR / \"speakers.json\"\n",
        "feature_dir = DATA_DIR / \"features\"\n",
        "stats_path = DATA_DIR / \"feature_stats.json\"\n",
        "\n",
        "def get_latest_checkpoint(checkpoint_dir):\n",
        "    output_path = None\n",
        "    for checkpoint_path in checkpoint_dir.iterdir():\n",
        "        checkpoint_name = str(checkpoint_path.name)\n",
        "        if checkpoint_name.endswith(\".nemo\"):\n",
        "            output_path = checkpoint_path\n",
        "            break\n",
        "        if checkpoint_name.endswith(\"last.ckpt\"):\n",
        "            output_path = checkpoint_path\n",
        "\n",
        "    if not output_path:\n",
        "        raise ValueError(f\"Could not find latest checkpoint in {checkpoint_dir}\")\n",
        "\n",
        "    return output_path\n",
        "\n",
        "# HiFi-GAN model for generating audio predictions from FastPitch output\n",
        "vocoder_type = \"hifigan\"\n",
        "vocoder_checkpoint_path = get_latest_checkpoint(hifigan_exp_output_dir / \"checkpoints\")\n",
        "\n",
        "run_id = \"test_run\"\n",
        "exp_dir = root_dir / \"exps\"\n",
        "fastpitch_exp_output_dir = exp_dir / \"FastPitch\" / run_id\n",
        "fastpitch_log_dir = fastpitch_exp_output_dir / \"logs\"\n",
        "\n",
        "if torch.cuda.is_available():\n",
        "    accelerator=\"gpu\"\n",
        "    batch_size = 32\n",
        "else:\n",
        "    accelerator=\"cpu\"\n",
        "    batch_size = 4\n",
        "\n",
        "args = [\n",
        "    f\"--config-path={fastpitch_config_dir}\",\n",
        "    f\"--config-name={fastpitch_config_filename}\",\n",
        "    f\"n_speakers={num_speakers}\",\n",
        "    f\"speaker_path={speaker_path}\",\n",
        "    f\"max_epochs={epochs}\",\n",
        "    f\"weighted_sampling_steps_per_epoch={steps_per_epoch}\",\n",
        "    f\"phoneme_dict_path={phoneme_dict_path}\",\n",
        "    f\"heteronyms_path={heteronyms_path}\",\n",
        "    f\"feature_stats_path={stats_path}\",\n",
        "    f\"log_dir={fastpitch_log_dir}\",\n",
        "    f\"vocoder_type={vocoder_type}\",\n",
        "    f\"vocoder_checkpoint_path=\\\\'{vocoder_checkpoint_path}\\\\'\",\n",
        "    f\"trainer.accelerator={accelerator}\",\n",
        "    f\"exp_manager.exp_dir={exp_dir}\",\n",
        "    f\"+exp_manager.version={run_id}\",\n",
        "    f\"+train_ds_meta.{dataset_name}.manifest_path={train_manifest_filepath}\",\n",
        "    f\"+train_ds_meta.{dataset_name}.audio_dir={audio_dir}\",\n",
        "    f\"+train_ds_meta.{dataset_name}.feature_dir={feature_dir}\",\n",
        "    f\"+val_ds_meta.{dataset_name}.manifest_path={dev_manifest_filepath}\",\n",
        "    f\"+val_ds_meta.{dataset_name}.audio_dir={audio_dir}\",\n",
        "    f\"+val_ds_meta.{dataset_name}.feature_dir={feature_dir}\",\n",
        "    f\"+log_ds_meta.{dataset_name}.manifest_path={dev_manifest_filepath}\",\n",
        "    f\"+log_ds_meta.{dataset_name}.audio_dir={audio_dir}\",\n",
        "    f\"+log_ds_meta.{dataset_name}.feature_dir={feature_dir}\"\n",
        "]"
      ],
      "metadata": {
        "id": "8MdMXnOAIFvj"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "run_script(fastpitch_training_script, args)"
      ],
      "metadata": {
        "id": "apl7TvW0TaEG"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "During training, the model will automatically save spectrogram and audio predictions for all files specified in the `log_ds_meta` manifest."
      ],
      "metadata": {
        "id": "Z01Fq7WRl7Di"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "faspitch_log_epoch_dir = fastpitch_log_dir / \"epoch_10\" / dataset_name\n",
        "!ls $faspitch_log_epoch_dir"
      ],
      "metadata": {
        "id": "E8rVKnKN5HDa"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "This makes it easy to listen to the audio to determine how well the model is performing. We can decide to stop training when either:\n",
        "\n",
        "*   The predicted audio stops improving in between epochs.\n",
        "*   The predicted spectrogram stops changing in between epochs.\n",
        "\n",
        "**Note that the dataset in this tutorial is too small to get good quality audio output.**"
      ],
      "metadata": {
        "id": "PeNaxoCzN7Ii"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "audio_filepath = faspitch_log_epoch_dir / \"p225_143.wav\"\n",
        "spectrogram_filepath = faspitch_log_epoch_dir / \"p225_143_spec.png\"\n",
        "\n",
        "ipd.display(ipd.Audio(audio_filepath))\n",
        "ipd.display(ipd.Image(spectrogram_filepath))"
      ],
      "metadata": {
        "id": "ynZdcnKc3CRF"
      },
      "execution_count": null,
      "outputs": []
    }
  ]
}