File size: 55,117 Bytes
7934b29 |
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 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 |
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 1000
},
"colab_type": "code",
"id": "cvXwyS263AMk",
"outputId": "57646cdd-58c1-4ddb-8805-9178cb0a2048"
},
"outputs": [],
"source": [
"\"\"\"\n",
"You can run either this notebook locally (if you have all the dependencies and a GPU) or on Google Colab.\n",
"\n",
"Instructions for setting up Colab are as follows:\n",
"1. Open a new Python 3 notebook.\n",
"2. Import this notebook from GitHub (File -> Upload Notebook -> \"GITHUB\" tab -> copy/paste GitHub URL)\n",
"3. Connect to an instance with a GPU (Runtime -> Change runtime type -> select \"GPU\" for hardware accelerator)\n",
"4. Run this cell to set up dependencies.\n",
"\"\"\"\n",
"# If you're using Google Colab and not running locally, run this cell.\n",
"\n",
"## Install dependencies\n",
"!pip install wget\n",
"!apt-get install sox libsndfile1 ffmpeg\n",
"!pip install text-unidecode\n",
"\n",
"# ## Install NeMo\n",
"BRANCH = 'r1.17.0'\n",
"!python -m pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[asr]\n",
"\n",
"## Install TorchAudio\n",
"!pip install torchaudio>=0.13.0 -f https://download.pytorch.org/whl/torch_stable.html\n",
"\n",
"## Grab the config we'll use in this example\n",
"!mkdir configs"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "Kqg4Rwki4jBX"
},
"source": [
"# Introduction\n",
"\n",
"Data augmentation is a useful method to improve the performance of models which is applicable across multiple domains. Certain augmentations can also substantially improve robustness of models to noisy samples. \n",
"\n",
"In this notebook, we describe how to construct an augmentation pipeline inside [Neural Modules (NeMo)](https://github.com/NVIDIA/NeMo), enable augmented training of a [MatchboxNet model](https://arxiv.org/abs/2004.08531 ) ( based on QuartzNet, from the paper [\"QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions\"](https://arxiv.org/abs/1910.10261)) and finally how to construct custom augmentations to add to NeMo.\n",
"\n",
"The notebook will follow the steps below:\n",
"\n",
" - Dataset preparation: Preparing a noise dataset using an example file.\n",
"\n",
" - Construct a data augmentation pipeline.\n",
" \n",
" - Construct a custom augmentation and register it for use in NeMo."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "5XieMEo84pJ-"
},
"source": [
"## Note\n",
"Data augmentation is valuable for many datasets, but it comes at the cost of increased training time if samples are augmented during training time. Certain augmentations are particularly costly, in terms of how much time they take to process a single sample. A few examples of slow augmentations available in NeMo are : \n",
"\n",
" - Speed Perturbation\n",
" - Time Stretch Perturbation (Sample level)\n",
" - Noise Perturbation\n",
" - Impulse Perturbation\n",
" - Time Stretch Augmentation (Batch level, Neural Module)\n",
" \n",
"For such augmentations, it is advisable to pre-process the dataset offline for a one time preprocessing cost and then train the dataset on this augmented training set."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "Tgc_ZHDl4sMy"
},
"source": [
"## Taking a Look at Our Data (AN4)\n",
"\n",
"The AN4 dataset, also known as the Alphanumeric dataset, was collected and published by Carnegie Mellon University. It consists of recordings of people spelling out addresses, names, telephone numbers, etc., one letter or number at a time, as well as their corresponding transcripts. We choose to use AN4 for this tutorial because it is relatively small, with 948 training and 130 test utterances, and so it trains quickly.\n",
"\n",
"Before we get started, let's download and prepare the dataset. The utterances are available as `.sph` files, so we will need to convert them to `.wav` for later processing. Please make sure you have [Sox](http://sox.sourceforge.net/) installed for this step (instructions to setup your environment are available at the beginning of this notebook)."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "DtLm_XuQ3pmk"
},
"outputs": [],
"source": [
"# This is where the an4/ directory will be placed.\n",
"# Change this if you don't want the data to be extracted in the current directory.\n",
"data_dir = '.'"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 102
},
"colab_type": "code",
"id": "HjfLhUtH4wNc",
"outputId": "f0a9cd46-6709-49dd-9103-1e0ef61de745"
},
"outputs": [],
"source": [
"import glob\n",
"import os\n",
"import subprocess\n",
"import tarfile\n",
"import wget\n",
"\n",
"# Download the dataset. This will take a few moments...\n",
"print(\"******\")\n",
"if not os.path.exists(data_dir + '/an4_sphere.tar.gz'):\n",
" an4_url = 'https://dldata-public.s3.us-east-2.amazonaws.com/an4_sphere.tar.gz' # for the original source, please visit http://www.speech.cs.cmu.edu/databases/an4/an4_sphere.tar.gz \n",
" an4_path = wget.download(an4_url, data_dir)\n",
" print(f\"Dataset downloaded at: {an4_path}\")\n",
"else:\n",
" print(\"Tarfile already exists.\")\n",
" an4_path = data_dir + '/an4_sphere.tar.gz'\n",
"\n",
"if not os.path.exists(data_dir + '/an4/'):\n",
" # Untar and convert .sph to .wav (using sox)\n",
" tar = tarfile.open(an4_path)\n",
" tar.extractall(path=data_dir)\n",
"\n",
" print(\"Converting .sph to .wav...\")\n",
" sph_list = glob.glob(data_dir + '/an4/**/*.sph', recursive=True)\n",
" for sph_path in sph_list:\n",
" wav_path = sph_path[:-4] + '.wav'\n",
" cmd = [\"sox\", sph_path, wav_path]\n",
" subprocess.run(cmd)\n",
"print(\"Finished conversion.\\n******\")\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "HqJmf4WB5P1x"
},
"source": [
"You should now have a folder called `an4` that contains `etc/an4_train.transcription`, `etc/an4_test.transcription`, audio files in `wav/an4_clstk` and `wav/an4test_clstk`, along with some other files we will not need.\n",
"\n",
"We now build a few manifest files which will be used later:"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 85
},
"colab_type": "code",
"id": "AmR6CH025C8E",
"outputId": "0cd776ea-078f-4ab8-8a79-eed3e1c05839"
},
"outputs": [],
"source": [
"# --- Building Manifest Files --- #\n",
"import json\n",
"import librosa\n",
"\n",
"# Function to build a manifest\n",
"def build_manifest(transcripts_path, manifest_path, wav_path):\n",
" with open(transcripts_path, 'r') as fin:\n",
" with open(manifest_path, 'w') as fout:\n",
" for line in fin:\n",
" # Lines look like this:\n",
" # <s> transcript </s> (fileID)\n",
" transcript = line[: line.find('(')-1].lower()\n",
" transcript = transcript.replace('<s>', '').replace('</s>', '')\n",
" transcript = transcript.strip()\n",
"\n",
" file_id = line[line.find('(')+1 : -2] # e.g. \"cen4-fash-b\"\n",
" audio_path = os.path.join(\n",
" data_dir, wav_path,\n",
" file_id[file_id.find('-')+1 : file_id.rfind('-')],\n",
" file_id + '.wav')\n",
"\n",
" duration = librosa.core.get_duration(filename=audio_path)\n",
"\n",
" # Write the metadata to the manifest\n",
" metadata = {\n",
" \"audio_filepath\": audio_path,\n",
" \"duration\": duration,\n",
" \"text\": transcript\n",
" }\n",
" json.dump(metadata, fout)\n",
" fout.write('\\n')\n",
" \n",
"# Building Manifests\n",
"print(\"******\")\n",
"train_transcripts = data_dir + '/an4/etc/an4_train.transcription'\n",
"train_manifest = data_dir + '/an4/train_manifest.json'\n",
"build_manifest(train_transcripts, train_manifest, 'an4/wav/an4_clstk')\n",
"print(\"Training manifest created.\")\n",
"\n",
"test_transcripts = data_dir + '/an4/etc/an4_test.transcription'\n",
"test_manifest = data_dir + '/an4/test_manifest.json'\n",
"build_manifest(test_transcripts, test_manifest, 'an4/wav/an4test_clstk')\n",
"print(\"Test manifest created.\")\n",
"print(\"******\")"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "EQsXzh7x5zIQ"
},
"source": [
"## Prepare the path to manifest files"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "vmOa0IRC5eW4"
},
"outputs": [],
"source": [
"dataset_basedir = os.path.join(data_dir, 'an4')\n",
"\n",
"train_dataset = os.path.join(dataset_basedir, 'train_manifest.json')\n",
"test_dataset = os.path.join(dataset_basedir, 'test_manifest.json')"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "pz9LC3yZ6J1Q"
},
"source": [
"## Read a few rows of the manifest file \n",
"\n",
"Manifest files are the data structure used by NeMo to declare a few important details about the data :\n",
"\n",
"1) `audio_filepath`: Refers to the path to the raw audio file <br>\n",
"2) `text`: The text transcript of this sample <br>\n",
"3) `duration`: The length of the audio file, in seconds."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "3OzZQiX751iz"
},
"outputs": [],
"source": [
"!head -n 5 {train_dataset}"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "pD9bprV66Oai"
},
"source": [
"# Data Augmentation Pipeline\n",
"\n",
"Constructing a data augmentation pipeline in NeMo is as simple as composing a nested dictionary that describes two things :\n",
"\n",
"1) The probability of that augmentation occurring - using the `prob` keyword <br>\n",
"2) The keyword arguments required by that augmentation class\n",
"\n",
"Below, we show a few samples of these augmentations. Note, in order to distinguish between the original sample and the perturbed sample, we exaggerate the perturbation strength significantly."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "l5bc7gYO6MHG"
},
"outputs": [],
"source": [
"import torch\n",
"import IPython.display as ipd"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "L8Bd8s3e6TeK"
},
"source": [
"## Audio file preparation "
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "g7f9riZz6Qnj"
},
"outputs": [],
"source": [
"# Import the data augmentation component from ASR collection\n",
"from nemo.collections.asr.parts.preprocessing import perturb, segment"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "wK8uwpt16d6I"
},
"outputs": [],
"source": [
"# Lets see the available perturbations\n",
"perturb.perturbation_types"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "IP1VpkOA6nE-"
},
"source": [
"### Obtain a baseline audio file"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "sj4DNMmZ6ktm"
},
"outputs": [],
"source": [
"filepath = librosa.util.example(key='vibeace', hq=True)\n",
"sample, sr = librosa.core.load(filepath)\n",
"\n",
"ipd.Audio(sample, rate=sr)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "M9mZNm296tNf"
},
"source": [
"### Convert to WAV format"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "QDjlgLc-6vtq"
},
"outputs": [],
"source": [
"import soundfile as sf\n",
"\n",
"# lets convert this ogg file into a wave to be compatible with NeMo\n",
"if not os.path.exists('./media'):\n",
" os.makedirs('./media/')\n",
" \n",
"filename = 'Kevin_MacLeod_-_Vibe_Ace.wav'\n",
"filepath = os.path.join('media', filename)\n",
"\n",
"sf.write(filepath, sample, samplerate=sr)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "FEkV-ikT6xgB"
},
"outputs": [],
"source": [
"sample, sr = librosa.core.load(filepath)\n",
"ipd.Audio(sample, rate=sr)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "gmuwEwIQ6zK3"
},
"outputs": [],
"source": [
"# NeMo has its own support class for loading wav files\n",
"def load_audio() -> segment.AudioSegment:\n",
" filename = 'Kevin_MacLeod_-_Vibe_Ace.wav'\n",
" filepath = os.path.join('media', filename)\n",
" sample_segment = segment.AudioSegment.from_file(filepath, target_sr=sr)\n",
" return sample_segment\n",
"\n",
"sample_segment = load_audio()\n",
"ipd.Audio(sample_segment.samples, rate=sr)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "hTnf1g1y63wZ"
},
"source": [
"## White Noise Perturbation\n",
"\n",
"White Noise perturbation is performed by the following steps : <br>\n",
"1) Randomly sample the amplitude of the noise from a uniformly distributed range (defined in dB) <br>\n",
"2) Sample gaussian noise (mean = 0, std = 1) with same length as audio signal <br>\n",
"3) Scale this gaussian noise by the amplitude (in dB scale) <br>\n",
"4) Add this noise vector to the original sample\n",
"\n",
"Notably, the original signal should not have a \"hissing sound\" constantly present in the perturbed version."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "2jaPQyUY65ij"
},
"outputs": [],
"source": [
"white_noise = perturb.WhiteNoisePerturbation(min_level=-50, max_level=-30)\n",
"\n",
"# Perturb the audio file\n",
"sample_segment = load_audio()\n",
"white_noise.perturb(sample_segment)\n",
"\n",
"ipd.Audio(sample_segment.samples, rate=sr)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "2dfwesJU7DhN"
},
"source": [
"## Shift Perturbation\n",
"\n",
"Shift perturbation is performed by the following steps : <br>\n",
"1) Randomly sample the shift factor of the signal from a uniformly distributed range (defined in milliseconds) <br>\n",
"2) Depending on the sign of the shift, we shift the original signal to the left or the right. <br>\n",
"3) The boundary locations are filled with zeros after the shift of the signal <br>\n",
"\n",
"Notably, the perturbed signal below skips the first 25 to 50 seconds of the original audio below, and the remainder of the time is simply silence. "
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "2ONq8dBI7BZf"
},
"outputs": [],
"source": [
"shift = perturb.ShiftPerturbation(min_shift_ms=25000.0, max_shift_ms=50000.0)\n",
"\n",
"# Perturb the audio file \n",
"sample_segment = load_audio()\n",
"shift.perturb(sample_segment)\n",
"\n",
"ipd.Audio(sample_segment.samples, rate=sr)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "kywA3h4T7G_S"
},
"source": [
"## Data Dependent Perturbations\n",
"\n",
"Some perturbations require an external data source in order to perturb the original sample. Noise Perturbation is a perfect example of one such augmentation that requires an external noise source dataset in order to perturb the original data."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "eYm2DgGQ7KPe"
},
"source": [
"### Preparing a manifest of \"noise\" samples"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "RXZ1o85E7FLT"
},
"outputs": [],
"source": [
"# Lets prepare a manifest file using the baseline file itself, cut into 1 second segments\n",
"\n",
"def write_manifest(filepath, data_dir='./media/', manifest_name='noise_manifest', duration_max=None, duration_stride=1.0, filter_long=False, duration_limit=10.0):\n",
" if duration_max is None:\n",
" duration_max = 1e9\n",
" \n",
" with open(os.path.join(data_dir, manifest_name + '.json'), 'w') as fout:\n",
" \n",
" try:\n",
" x, _sr = librosa.load(filepath)\n",
" duration = librosa.get_duration(x, sr=_sr)\n",
"\n",
" except Exception:\n",
" print(f\"\\n>>>>>>>>> WARNING: Librosa failed to load file {filepath}. Skipping this file !\\n\")\n",
" return\n",
"\n",
" if filter_long and duration > duration_limit:\n",
" print(f\"Skipping sound sample {filepath}, exceeds duration limit of {duration_limit}\")\n",
" return\n",
"\n",
" offsets = []\n",
" durations = []\n",
"\n",
" if duration > duration_max:\n",
" current_offset = 0.0\n",
"\n",
" while current_offset < duration:\n",
" difference = duration - current_offset\n",
" segment_duration = min(duration_max, difference)\n",
"\n",
" offsets.append(current_offset)\n",
" durations.append(segment_duration)\n",
"\n",
" current_offset += duration_stride\n",
"\n",
" else:\n",
" offsets.append(0.0)\n",
" durations.append(duration)\n",
"\n",
"\n",
" for duration, offset in zip(durations, offsets):\n",
" metadata = {\n",
" 'audio_filepath': filepath,\n",
" 'duration': duration,\n",
" 'label': 'noise',\n",
" 'text': '_', # for compatibility with ASRAudioText collection\n",
" 'offset': offset,\n",
" }\n",
"\n",
" json.dump(metadata, fout)\n",
" fout.write('\\n')\n",
" fout.flush()\n",
"\n",
" print(f\"Wrote {len(durations)} segments for filename {filename}\")\n",
" \n",
" print(\"Finished preparing manifest !\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "wLTT8jlP7NdU"
},
"outputs": [],
"source": [
"filename = 'Kevin_MacLeod_-_Vibe_Ace.wav'\n",
"filepath = os.path.join('media', filename)\n",
"\n",
"# Write a \"noise\" manifest file\n",
"write_manifest(filepath, manifest_name='noise_1s', duration_max=1.0, duration_stride=1.0)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "izbdrSmd7PY5"
},
"outputs": [],
"source": [
"# Lets read this noise manifest file\n",
"noise_manifest_path = os.path.join('media', 'noise_1s.json')\n",
"\n",
"!head -n 5 {noise_manifest_path}"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "82yq0TOV7Q_4"
},
"outputs": [],
"source": [
"# Lets create a helper method to load the first file in the train dataset of AN4\n",
"# Load the first sample in the manifest\n",
"def load_gsc_sample() -> segment.AudioSegment:\n",
" with open(train_dataset, 'r') as f:\n",
" line = f.readline()\n",
" \n",
" line = json.loads(line)\n",
" gsc_filepath = line['audio_filepath']\n",
" sample_segment = segment.AudioSegment.from_file(gsc_filepath)\n",
" return sample_segment\n",
"\n",
"gsc_sample_segment = load_gsc_sample()\n",
"ipd.Audio(gsc_sample_segment.samples, rate=16000)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "zV9ypBqz7V9a"
},
"source": [
"## Noise Augmentation\n",
"\n",
"Noise perturbation is performed by the following steps : <br>\n",
"1) Randomly sample the amplitude scale of the noise sample from a uniformly distributed range (defined in dB) <br>\n",
"2) Randomly choose an audio clip from the set of noise audio samples available <br>\n",
"3) Compute the gain (in dB) required for the noise clip as compared to the original sample and scale the noise by this factor <br>\n",
"4) If the noise snippet is of shorter duration than the original audio, then randomly select an index in time from the original sample, where the noise snippet will be added <br>\n",
"5) If instead the noise snippet is longer than the duration of the original audio, then randomly subsegment the noise snippet and add the full snippet to the original audio <br>\n",
"\n",
"Notably, the noise perturbed sample should sound as if there are two sounds playing at the same time (overlapping audio) as compared to the original signal. The magnitude of the noise will be dependent on step (3) and the location where the noise is added will depend on steps (4) and (5)."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "cjSXci1v7Tlg"
},
"outputs": [],
"source": [
"import random\n",
"rng = 42 #note you can use integer to be as random seed to reproduce result \n",
"noise = perturb.NoisePerturbation(manifest_path=noise_manifest_path,\n",
" min_snr_db=-10, max_snr_db=-10,\n",
" max_gain_db=300.0, rng=rng)\n",
"\n",
"# Perturb the audio file \n",
"sample_segment = load_gsc_sample()\n",
"noise.perturb(sample_segment)\n",
"\n",
"ipd.Audio(sample_segment.samples, rate=16000)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## RIR and Noise Perturbation\n",
"RIR augmentation with additive foreground and background noise.\n",
"In this implementation audio data is augmented by first convolving the audio with a Room Impulse Response\n",
"and then adding foreground noise and background noise at various SNRs. RIR, foreground and background noises\n",
"should either be supplied with a manifest file or as tarred audio files (faster)."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Prepare rir data and manifest"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# This is where the rir data will be downloaded.\n",
"# Change this if you don't want the data to be extracted in the current directory.\n",
"rir_data_path = '.'\n",
"!python ../../scripts/dataset_processing/get_openslr_rir_data.py --data_root {rir_data_path}\n",
"rir_manifest_path = os.path.join(rir_data_path, 'processed', 'rir.json')\n",
"!head -n 3 {rir_manifest_path}"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Create RIR instance"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"rir = perturb.RirAndNoisePerturbation(rir_manifest_path=rir_manifest_path, \n",
" rir_prob=1,\n",
" noise_manifest_paths=[noise_manifest_path], # use noise_manifest_path from previous step\n",
" bg_noise_manifest_paths=[noise_manifest_path],\n",
" min_snr_db=[20], # foreground noise snr\n",
" max_snr_db=[20],\n",
" bg_min_snr_db=[20], # background noise snr\n",
" bg_max_snr_db=[20],\n",
" noise_tar_filepaths=[None], # `[None]` to indicates that noise audio files are not tar.\n",
" bg_noise_tar_filepaths=[None])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Perturb the audio"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"sample_segment = load_gsc_sample()\n",
"rir.perturb(sample_segment)\n",
"ipd.Audio(sample_segment.samples, rate=16000)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "kJjUkGJu7ern"
},
"source": [
"## Speed Perturbation\n",
"\n",
"Speed perturbation changes the speed of the speech, but does not preserve pitch of the sound. Try a few random augmentations to see how the pitch changes with change in duration of the audio file.\n",
"\n",
"**Note**: This is a very slow augmentation and is not advised to perform online augmentation for large datasets as it can dramatically increase training time."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "Ic-ziInU7ZKC"
},
"outputs": [],
"source": [
"resample_type = 'kaiser_best' # Can be ['kaiser_best', 'kaiser_fast', 'fft', 'scipy']\n",
"speed = perturb.SpeedPerturbation(sr, resample_type, min_speed_rate=0.5, max_speed_rate=2.0, num_rates=-1)\n",
"\n",
"# Perturb the audio file \n",
"sample_segment = load_gsc_sample()\n",
"speed.perturb(sample_segment)\n",
"\n",
"ipd.Audio(sample_segment.samples, rate=16000)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "bhHX3dyh7jPq"
},
"source": [
"## Time Stretch Perturbation\n",
"\n",
"Time Stretch perturbation changes the speed of the speech, and also preserve pitch of the sound. \n",
"Try a few random augmentations to see how the pitch remains close to the same with change in duration of the audio file."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "8_kNSfcK7lfP"
},
"source": [
"### Note about speed optimizations\n",
"\n",
"Time stretch is a costly augmentation, and can easily cause training time to increase drastically. It is suggested that one installs the `numba` library using conda to use a more optimized augmentation kernel.\n",
"\n",
"```python\n",
"conda install numba\n",
"```"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "Dpeb0QUZ7g3l"
},
"outputs": [],
"source": [
"time_stretch = perturb.TimeStretchPerturbation(min_speed_rate=0.5, max_speed_rate=2.0, num_rates=3)\n",
"\n",
"# Perturb the audio file \n",
"sample_segment = load_gsc_sample()\n",
"time_stretch.perturb(sample_segment)\n",
"\n",
"ipd.Audio(sample_segment.samples, rate=16000)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "vhH1-Ga87rCX"
},
"source": [
"# Augmentation Pipeline\n",
"\n",
"The augmentation pipeline can be constructed in multiple ways, either explicitly by instantiating the objects of these perturbations or implicitly by providing the arguments to these augmentations as a nested dictionary.\n",
"\n",
"We will show both approaches in the following sections"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "RC8_NOD97tlW"
},
"source": [
"## Explicit definition"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "UwWE7swo72WP"
},
"source": [
"### Instantiate the perturbations"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "GdLYn0hx7pRU"
},
"outputs": [],
"source": [
"perturbations = [\n",
" perturb.WhiteNoisePerturbation(min_level=-90, max_level=-46),\n",
" perturb.GainPerturbation(min_gain_dbfs=0, max_gain_dbfs=50),\n",
" perturb.NoisePerturbation(manifest_path=noise_manifest_path,\n",
" min_snr_db=0, max_snr_db=50, max_gain_db=300.0)\n",
"]"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "CDSSbZ8w7zzR"
},
"source": [
"### Select chance of perturbations being applied"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "NmoxfLSL7xPJ"
},
"outputs": [],
"source": [
"probas = [1.0, 1.0, 0.5]"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "wl0tnrMq79Jh"
},
"source": [
"### Prepare the audio augmentation object"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "nO6T4U4f767o"
},
"outputs": [],
"source": [
"augmentations = list(zip(probas, perturbations))\n",
"\n",
"audio_augmentations = perturb.AudioAugmentor(augmentations)\n",
"audio_augmentations._pipeline"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "9cgI9yUx8Cyv"
},
"source": [
"## Implicit definition\n",
"\n",
"Implicit definitions are preferred since they can be prepared in the actual configuration object."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "tiqrKFTM7_mH"
},
"outputs": [],
"source": [
"perturb.perturbation_types # Available perturbations"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "dbeXwLdw8VEc"
},
"source": [
"### Prepare the nested dictionary"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "mbE0qEA98TRI"
},
"outputs": [],
"source": [
"audio_augmentations = dict(\n",
" white_noise = dict(\n",
" prob=1.0,\n",
" min_level=-90,\n",
" max_level=-46\n",
" ),\n",
" gain = dict(\n",
" prob=1.0,\n",
" min_gain_dbfs=0,\n",
" max_gain_dbfs=50\n",
" ),\n",
" noise = dict(\n",
" prob=0.5,\n",
" manifest_path=noise_manifest_path,\n",
" min_snr_db=0,\n",
" max_snr_db=50,\n",
" max_gain_db=300.0\n",
" )\n",
")\n",
"\n",
"audio_augmentations"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "tcsoCe9-8ZM9"
},
"source": [
"### Supply `augmentor` as an argument to the `model.train_ds` config\n",
"\n",
"Most of the common datasets used by ASR models support the keyword `augmentor` - which can include a nested dictionary defining the implicit definition of an augmentation pipeline.\n",
"\n",
"Note, all ASR models support implicit declaration of augmentations. This includes - \n",
"\n",
"1) Speech To Label Models <br>\n",
"2) Speech To Text Models <br>\n",
"3) Speech To Text Models with BPE/WPE Support<br>"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "0WOJC0fdBL5J"
},
"source": [
"# Training - Application of augmentations\n",
"\n",
"We will be describing the data loaders for a MatchboxNet model from the paper \"[MatchboxNet: 1D Time-Channel Separable Convolutional Neural Network Architecture for Speech Commands Recognition](https://arxiv.org/abs/2004.08531)\". The benefit of MatchboxNet over JASPER models is that they use Separable Convolutions, which greatly reduce the number of parameters required to get good model accuracy.\n",
"\n",
"<ins>Care must be taken not to apply augmentations to the test set!</ins>\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "7iDWiIrzBzUA"
},
"outputs": [],
"source": [
"from omegaconf import OmegaConf"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "yv3KWNjcAUnQ"
},
"outputs": [],
"source": [
"# We will download the MatchboxNet configuration file for either v1 or v2 dataset here\n",
"DATASET_VER = 1\n",
"\n",
"if DATASET_VER == 1:\n",
" MODEL_CONFIG = \"matchboxnet_3x1x64_v1.yaml\"\n",
"else:\n",
" MODEL_CONFIG = \"matchboxnet_3x1x64_v2.yaml\"\n",
"\n",
"if not os.path.exists(f\"configs/{MODEL_CONFIG}\"):\n",
" !wget -P configs/ \"https://raw.githubusercontent.com/NVIDIA/NeMo/$BRANCH/examples/asr/conf/matchboxnet/{MODEL_CONFIG}\""
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "vOcv0ri3BkmA"
},
"outputs": [],
"source": [
"# This line will load the entire config of the MatchboxNet model\n",
"config_path = f\"configs/{MODEL_CONFIG}\"\n",
"config = OmegaConf.load(config_path)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "mLsyceMSCIHV"
},
"source": [
"### Augmentation in train set only\n",
"\n",
"Note how the train dataset config supports the `augmentor` implicit definition, however the test config does not.\n",
"\n",
"This is essential to avoid mistakenly performing Test Time Augmentation."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "VgUVm7lGB8Cz"
},
"outputs": [],
"source": [
"# Has `augmentor`\n",
"print(OmegaConf.to_yaml(config.model.train_ds))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "gURwQ2eyCE7o"
},
"outputs": [],
"source": [
"# Does not have `augmentor`\n",
"print(OmegaConf.to_yaml(config.model.test_ds))"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "_UV74AVlCo_m"
},
"source": [
"# Custom Perturbations\n",
"\n",
"We can define and use custom perturbations as required simply by extending the `Perturbation` class. \n",
"\n",
"Let's look at how we can build a custom Noise Perturbation that we can use to evaluate the effect of noise at inference time, in order to measure the model's robustness to noise\n",
"\n",
"In evaluation mode, we want to set an explicit value for the `snr_db` parameter instead of uniformly sample it from a range. This allows us to control the signal to noise ratio without relying on randomness from the training implementation of `NoisePerturbation`.\n",
"\n",
"Further, we force a random seed in order to produce reproduceable results on the evaluation set.\n",
"\n",
"With this combination, we can easily evaluate each sample in the test set `S` times (`S` being the number of random seeds), and can evaluate each of these samples at `D` levels of Signal to Noise Ratio (in dB). "
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "Q9YBmBiZCbAX"
},
"outputs": [],
"source": [
"# We use a NeMo utility to parse the manifest file for us\n",
"from nemo.collections.common.parts.preprocessing import collections, parsers\n",
"\n",
"class NoisePerturbationEval(perturb.Perturbation):\n",
" def __init__(\n",
" self, manifest_path=None, snr_db=40, max_gain_db=300.0, rng=None,\n",
" ):\n",
" self._manifest = collections.ASRAudioText(manifest_path, parser=parsers.make_parser([]))\n",
" self._snr_db = snr_db\n",
" self._max_gain_db = max_gain_db\n",
" random.seed(rng) if rng else None\n",
" \n",
" # This is mostly obtained from the original NoisePerturbation class itself\n",
" def perturb(self, data):\n",
" snr_db = self._snr_db\n",
" noise_record = random.sample(self._manifest.data, 1)[0]\n",
" noise = AudioSegment.from_file(noise_record.audio_file, target_sr=data.sample_rate)\n",
" noise_gain_db = min(data.rms_db - noise.rms_db - snr_db, self._max_gain_db)\n",
"\n",
" # calculate noise segment to use\n",
" start_time = 0.0\n",
" if noise.duration > (start_time + data.duration):\n",
" noise.subsegment(start_time=start_time, end_time=start_time + data.duration)\n",
"\n",
" # adjust gain for snr purposes and superimpose\n",
" noise.gain_db(noise_gain_db)\n",
"\n",
" if noise._samples.shape[0] < data._samples.shape[0]:\n",
" noise_idx = data._samples.shape[0] // 2 # midpoint of audio\n",
" while (noise_idx + noise._samples.shape[0]) > data._samples.shape[0]:\n",
" noise_idx = noise_idx // 2 # half the initial starting point\n",
"\n",
" data._samples[noise_idx: noise_idx + noise._samples.shape[0]] += noise._samples\n",
"\n",
" else:\n",
" data._samples += noise._samples\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "qR8qiwSkC1eE"
},
"source": [
"## Registering augmentations\n",
"\n",
"We can use either approach to submit this test time augmentation to the Data Loaders.\n",
"\n",
"In order to obtain the convenience of the implicit method, we must register this augmentation into NeMo's directory of available augmentations. This can be done as follows -"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "40Z4Fm88CxWA"
},
"outputs": [],
"source": [
"perturb.register_perturbation(name='noise_eval', perturbation=NoisePerturbationEval)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "jVVbRxb-C4hB"
},
"outputs": [],
"source": [
"# Lets check the registry of allowed perturbations !\n",
"perturb.perturbation_types"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "2fiHz6CdC-B1"
},
"source": [
"## Overriding pre-existing augmentations\n",
"\n",
"**Note**: It is not allowed to overwrite already registered perturbations using the `perturb.register_perturbation` method. It will raise a `ValueError` in order to prevent overwriting the pre-existing perturbation types"
]
}
],
"metadata": {
"accelerator": "GPU",
"colab": {
"collapsed_sections": [],
"name": "Online_Noise_Augmentation.ipynb",
"provenance": [],
"toc_visible": true
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.7"
},
"pycharm": {
"stem_cell": {
"cell_type": "raw",
"metadata": {
"collapsed": false
},
"source": []
}
}
},
"nbformat": 4,
"nbformat_minor": 4
}
|