File size: 21,859 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 |
{
"nbformat": 4,
"nbformat_minor": 0,
"metadata": {
"colab": {
"name": "Offline_ASR.ipynb",
"provenance": [],
"collapsed_sections": [],
"toc_visible": true
},
"kernelspec": {
"name": "python3",
"display_name": "Python 3"
},
"accelerator": "GPU"
},
"cells": [
{
"cell_type": "markdown",
"metadata": {
"id": "_wIWPxBVc3_O"
},
"source": [
"# NeMo offline ASR\n",
"\n",
"This notebook demonstrates how to \n",
"\n",
"* transcribe an audio file (offline ASR) with greedy decoder\n",
"* extract timestamps information from the model to split audio into separate words\n",
"* use beam search decoder with N-gram language model re-scoring\n",
"\n",
"You may find more info on how to train and use language models for ASR models here:\n",
"https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "gzcsqceVdtj3"
},
"source": [
"## Installation\n",
"NeMo can be installed via simple pip command. \n",
"\n",
"Optional CTC beam search decoder might require restart of Colab runtime after installation."
]
},
{
"cell_type": "code",
"metadata": {
"id": "I9eIxAyKHREB"
},
"source": [
"BRANCH = 'r1.17.0'\n",
"try:\n",
" # Import NeMo Speech Recognition collection\n",
" import nemo.collections.asr as nemo_asr\n",
"except ModuleNotFoundError:\n",
" !python -m pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[all]\n",
"\n",
"# check if we have optional Plotly for visualization\n",
"try:\n",
" from plotly import graph_objects as go\n",
"except ModuleNotFoundError:\n",
" !pip install plotly\n",
" from plotly import graph_objects as go\n",
"\n",
"# check if we have optional ipywidgets for tqdm/notebook\n",
"try:\n",
" import ipywidgets\n",
"except ModuleNotFoundError:\n",
" !pip install ipywidgets\n",
"\n",
"# check if CTC beam decoders are installed\n",
"try:\n",
" import ctc_decoders\n",
"except ModuleNotFoundError:\n",
" # install beam search decoder\n",
" !apt-get update && apt-get install -y swig\n",
" !git clone https://github.com/NVIDIA/NeMo -b \"$BRANCH\"\n",
" !cd NeMo && bash scripts/asr_language_modeling/ngram_lm/install_beamsearch_decoders.sh\n",
" print('Restarting Colab runtime to successfully import built module.')\n",
" print('Please re-run the notebook.')\n",
" import os\n",
" os.kill(os.getpid(), 9)"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "code",
"metadata": {
"id": "-X2OyAxreGfl"
},
"source": [
"import numpy as np\n",
"# Import audio processing library\n",
"import librosa\n",
"# We'll use this to listen to audio\n",
"from IPython.display import Audio, display"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "zodyzdyTVXas"
},
"source": [
"## Instantiate pre-trained NeMo model\n",
"``from_pretrained(...)`` API downloads and initializes model directly from the cloud. \n",
"\n",
"Alternatively, ``restore_from(...)`` allows loading a model from a disk.\n",
"\n",
"To display available pre-trained models from the cloud, please use ``list_available_models()`` method."
]
},
{
"cell_type": "code",
"metadata": {
"id": "f_J9cuU1H6Bn"
},
"source": [
"nemo_asr.models.EncDecCTCModel.list_available_models()"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "x2LMVI9qqtEV"
},
"source": [
"Let's load a base English QuartzNet15x5 model."
]
},
{
"cell_type": "code",
"metadata": {
"id": "ZhWmR7lbvwSm"
},
"source": [
"asr_model = nemo_asr.models.EncDecCTCModel.from_pretrained(model_name='QuartzNet15x5Base-En', strict=False)"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "HESTZmIzzCEj"
},
"source": [
"## Get test audio clip"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "QPWn89l-zLXo"
},
"source": [
"Let's download and analyze a test audio signal."
]
},
{
"cell_type": "code",
"metadata": {
"id": "02gDfK7czSVV"
},
"source": [
"# Download audio sample which we'll try\n",
"# This is a sample from LibriSpeech dev clean subset - the model hasn't seen it before\n",
"AUDIO_FILENAME = '1919-142785-0028.wav'\n",
"!wget https://dldata-public.s3.us-east-2.amazonaws.com/1919-142785-0028.wav\n",
"\n",
"# load audio signal with librosa\n",
"signal, sample_rate = librosa.load(AUDIO_FILENAME, sr=None)\n",
"\n",
"# display audio player for the signal\n",
"display(Audio(data=signal, rate=sample_rate))\n",
"\n",
"# plot the signal in time domain\n",
"fig_signal = go.Figure(\n",
" go.Scatter(x=np.arange(signal.shape[0])/sample_rate,\n",
" y=signal, line={'color': 'green'},\n",
" name='Waveform',\n",
" hovertemplate='Time: %{x:.2f} s<br>Amplitude: %{y:.2f}<br><extra></extra>'),\n",
" layout={\n",
" 'height': 300,\n",
" 'xaxis': {'title': 'Time, s'},\n",
" 'yaxis': {'title': 'Amplitude'},\n",
" 'title': 'Audio Signal',\n",
" 'margin': dict(l=0, r=0, t=40, b=0, pad=0),\n",
" }\n",
")\n",
"fig_signal.show()\n",
"\n",
"# calculate amplitude spectrum\n",
"time_stride=0.01\n",
"hop_length = int(sample_rate*time_stride)\n",
"n_fft = 512\n",
"# linear scale spectrogram\n",
"s = librosa.stft(y=signal,\n",
" n_fft=n_fft,\n",
" hop_length=hop_length)\n",
"s_db = librosa.power_to_db(np.abs(s)**2, ref=np.max, top_db=100)\n",
"\n",
"# plot the signal in frequency domain\n",
"fig_spectrum = go.Figure(\n",
" go.Heatmap(z=s_db,\n",
" colorscale=[\n",
" [0, 'rgb(30,62,62)'],\n",
" [0.5, 'rgb(30,128,128)'],\n",
" [1, 'rgb(30,255,30)'],\n",
" ],\n",
" colorbar=dict(\n",
" ticksuffix=' dB'\n",
" ),\n",
" dx=time_stride, dy=sample_rate/n_fft/1000,\n",
" name='Spectrogram',\n",
" hovertemplate='Time: %{x:.2f} s<br>Frequency: %{y:.2f} kHz<br>Magnitude: %{z:.2f} dB<extra></extra>'),\n",
" layout={\n",
" 'height': 300,\n",
" 'xaxis': {'title': 'Time, s'},\n",
" 'yaxis': {'title': 'Frequency, kHz'},\n",
" 'title': 'Spectrogram',\n",
" 'margin': dict(l=0, r=0, t=40, b=0, pad=0),\n",
" }\n",
")\n",
"fig_spectrum.show()"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "jQSj-IhEhrtI"
},
"source": [
"## Offline inference\n",
"If we have an entire audio clip available, then we can do offline inference with a pre-trained model to transcribe it.\n",
"\n",
"The easiest way to do it is to call ASR model's ``transcribe(...)`` method that allows transcribing multiple files in a batch."
]
},
{
"cell_type": "code",
"metadata": {
"id": "s0ERrXIzKpwu"
},
"source": [
"# Convert our audio sample to text\n",
"files = [AUDIO_FILENAME]\n",
"transcript = asr_model.transcribe(paths2audio_files=files)[0]\n",
"print(f'Transcript: \"{transcript}\"')"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "_UOoj-WfQoL_"
},
"source": [
"## Extract timestamps and split words\n",
"``transcribe()`` generates a text applying a CTC greedy decoder to raw probabilities distribution over alphabet's characters from ASR model. We can get those raw probabilities with ``logprobs=True`` argument."
]
},
{
"cell_type": "code",
"metadata": {
"id": "-0Sk0C9-LmAR"
},
"source": [
"# softmax implementation in NumPy\n",
"def softmax(logits):\n",
" e = np.exp(logits - np.max(logits))\n",
" return e / e.sum(axis=-1).reshape([logits.shape[0], 1])\n",
"\n",
"# let's do inference once again but without decoder\n",
"logits = asr_model.transcribe(files, logprobs=True)[0]\n",
"probs = softmax(logits)\n",
"\n",
"# 20ms is duration of a timestep at output of the model\n",
"time_stride = 0.02\n",
"\n",
"# get model's alphabet\n",
"labels = list(asr_model.decoder.vocabulary) + ['blank']\n",
"labels[0] = 'space'\n",
"\n",
"# plot probability distribution over characters for each timestep\n",
"fig_probs = go.Figure(\n",
" go.Heatmap(z=probs.transpose(),\n",
" colorscale=[\n",
" [0, 'rgb(30,62,62)'],\n",
" [1, 'rgb(30,255,30)'],\n",
" ],\n",
" y=labels,\n",
" dx=time_stride,\n",
" name='Probs',\n",
" hovertemplate='Time: %{x:.2f} s<br>Character: %{y}<br>Probability: %{z:.2f}<extra></extra>'),\n",
" layout={\n",
" 'height': 300,\n",
" 'xaxis': {'title': 'Time, s'},\n",
" 'yaxis': {'title': 'Characters'},\n",
" 'title': 'Character Probabilities',\n",
" 'margin': dict(l=0, r=0, t=40, b=0, pad=0),\n",
" }\n",
")\n",
"fig_probs.show()"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "YiNMZBodIaSP"
},
"source": [
"It is easy to identify timesteps for space character."
]
},
{
"cell_type": "code",
"metadata": {
"id": "32aaW3HEJ89l"
},
"source": [
"# get timestamps for space symbols\n",
"spaces = []\n",
"\n",
"state = ''\n",
"idx_state = 0\n",
"\n",
"if np.argmax(probs[0]) == 0:\n",
" state = 'space'\n",
"\n",
"for idx in range(1, probs.shape[0]):\n",
" current_char_idx = np.argmax(probs[idx])\n",
" if state == 'space' and current_char_idx != 0 and current_char_idx != 28:\n",
" spaces.append([idx_state, idx-1])\n",
" state = ''\n",
" if state == '':\n",
" if current_char_idx == 0:\n",
" state = 'space'\n",
" idx_state = idx\n",
"\n",
"if state == 'space':\n",
" spaces.append([idx_state, len(pred)-1])"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "rqg4oxpsL8cW"
},
"source": [
"Then we can split original audio signal into separate words. It is worth to mention that all timestamps have a delay (or an offset) depending on the model. We need to take it into account for alignment."
]
},
{
"cell_type": "code",
"metadata": {
"id": "a-LSg9dSL_O1"
},
"source": [
"# calibration offset for timestamps: 180 ms\n",
"offset = -0.18\n",
"\n",
"# split the transcript into words\n",
"words = transcript.split()\n",
"\n",
"# cut words\n",
"pos_prev = 0\n",
"for j, spot in enumerate(spaces):\n",
" display(words[j])\n",
" pos_end = offset + (spot[0]+spot[1])/2*time_stride\n",
" display(Audio(signal[int(pos_prev*sample_rate):int(pos_end*sample_rate)],\n",
" rate=sample_rate))\n",
" pos_prev = pos_end\n",
"\n",
"display(words[j+1])\n",
"display(Audio(signal[int(pos_prev*sample_rate):],\n",
" rate=sample_rate))"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "Q8Jvwe4Ahncx"
},
"source": [
"## Offline inference with beam search decoder and N-gram language model re-scoring\n",
"\n",
"It is possible to use an external [KenLM](https://kheafield.com/code/kenlm/)-based N-gram language model to rescore multiple transcription candidates. \n",
"\n",
"Let's download and preprocess LibriSpeech 3-gram language model."
]
},
{
"cell_type": "code",
"metadata": {
"id": "EIh8wTVs5uH7"
},
"source": [
"import gzip\n",
"import os, shutil, wget\n",
"\n",
"lm_gzip_path = '3-gram.pruned.1e-7.arpa.gz'\n",
"if not os.path.exists(lm_gzip_path):\n",
" print('Downloading pruned 3-gram model.')\n",
" lm_url = 'http://www.openslr.org/resources/11/3-gram.pruned.1e-7.arpa.gz'\n",
" lm_gzip_path = wget.download(lm_url)\n",
" print('Downloaded the 3-gram language model.')\n",
"else:\n",
" print('Pruned .arpa.gz already exists.')\n",
"\n",
"uppercase_lm_path = '3-gram.pruned.1e-7.arpa'\n",
"if not os.path.exists(uppercase_lm_path):\n",
" with gzip.open(lm_gzip_path, 'rb') as f_zipped:\n",
" with open(uppercase_lm_path, 'wb') as f_unzipped:\n",
" shutil.copyfileobj(f_zipped, f_unzipped)\n",
" print('Unzipped the 3-gram language model.')\n",
"else:\n",
" print('Unzipped .arpa already exists.')\n",
"\n",
"lm_path = 'lowercase_3-gram.pruned.1e-7.arpa'\n",
"if not os.path.exists(lm_path):\n",
" with open(uppercase_lm_path, 'r') as f_upper:\n",
" with open(lm_path, 'w') as f_lower:\n",
" for line in f_upper:\n",
" f_lower.write(line.lower())\n",
"print('Converted language model file to lowercase.')"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "fLDbUkzzUAqW"
},
"source": [
"Let's instantiate ``BeamSearchDecoderWithLM`` module."
]
},
{
"cell_type": "code",
"metadata": {
"id": "_qgKa9L954bJ"
},
"source": [
"beam_search_lm = nemo_asr.modules.BeamSearchDecoderWithLM(\n",
" vocab=list(asr_model.decoder.vocabulary),\n",
" beam_width=16,\n",
" alpha=2, beta=1.5,\n",
" lm_path=lm_path,\n",
" num_cpus=max(os.cpu_count(), 1),\n",
" input_tensor=False)"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "NSH8EvL7USac"
},
"source": [
"Now we can check all transcription candidates along with their scores."
]
},
{
"cell_type": "code",
"metadata": {
"id": "nV1CAy0Dit-g"
},
"source": [
"beam_search_lm.forward(log_probs = np.expand_dims(probs, axis=0), log_probs_length=None)"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"# Greedy Decoding Time Stamps\n",
"\n",
"While the above approach works well for character based CTC models, it requires careful tuning of offset parameter as well as computation of the word time stamp offsets.\n",
"\n",
"We therefore provide a simple way to obtain greedy decoding word time stamps directly using the familiar \"model.transcribe()\" method, which works quite well for character and subword models.\n",
"\n",
"**Note**: We find that larger models that have converged to strong scores on the dataset usually have better word alignments. If evaluated on a completely out of domain audio sample, it might produce very poor time stamps."
],
"metadata": {
"id": "LPtMzLE4T7T-"
}
},
{
"cell_type": "code",
"source": [
"from omegaconf import OmegaConf, open_dict"
],
"metadata": {
"id": "z_0pO-TaUIHU"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"For the purposes of this demonstration, we will use Conformer CTC Large, a 120 M parameter model trained on thousands of hours of English speech."
],
"metadata": {
"id": "i0Epb8D-rW3-"
}
},
{
"cell_type": "code",
"source": [
"asr_model_subword = nemo_asr.models.ASRModel.from_pretrained(\"stt_en_conformer_ctc_large\")"
],
"metadata": {
"id": "Ky7OpuikbBTb"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"## CTC Decoding Strategy\n",
"\n",
"NeMo CTC models have an internal decoding strategy that can be updated after training. In our case, we will enable the greedy decoding step to compute word time stamps, as well as preserve the log probability predictions."
],
"metadata": {
"id": "vwN6wddTrhno"
}
},
{
"cell_type": "code",
"metadata": {
"id": "ubpcxp6z3ZF-"
},
"source": [
"decoding_cfg = asr_model_subword.cfg.decoding\n",
"print(OmegaConf.to_yaml(decoding_cfg))"
],
"execution_count": null,
"outputs": []
},
{
"cell_type": "code",
"source": [
"decoding_cfg.preserve_alignments = True\n",
"decoding_cfg.compute_timestamps = True\n",
"asr_model_subword.change_decoding_strategy(decoding_cfg)"
],
"metadata": {
"id": "pKUsMlUbUAxv"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"Next, we simply transcribe the audio file, and pass the flag `return_hypotheses=True`. This will return a list of `Hypothesis` objects instead of the predicted text."
],
"metadata": {
"id": "EdX0Drncr8Yl"
}
},
{
"cell_type": "code",
"source": [
"hypothesis = asr_model_subword.transcribe([AUDIO_FILENAME], return_hypotheses=True)[0]"
],
"metadata": {
"id": "SUkfIyYzUbaB"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "code",
"source": [
"print(\"Greedy prediction :\", hypothesis.text)"
],
"metadata": {
"id": "duaxOSPXUmQ0"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"## Hypothesis - Time Stamps\n",
"\n",
"Since we previously set the flag for `decoding_cfg.compute_timestamps`, the hypothesis now contains a dictionary in it, accessed via `hypothesis.timestep`. This dictionary contains multiple useful lists, detailing the time step at which some token was predicted, the character / subword / word time stamps."
],
"metadata": {
"id": "_5hfsiDGsM19"
}
},
{
"cell_type": "code",
"source": [
"timestamp_dict = hypothesis.timestep\n",
"print(\"Hypothesis contains following timestep information :\", list(timestamp_dict.keys()))"
],
"metadata": {
"id": "vh7K_9D1UrQp"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "code",
"source": [
"# 40ms is duration of a timestep at output of the Conformer\n",
"time_stride = 4 * asr_model_subword.cfg.preprocessor.window_stride\n",
"\n",
"##################################################################\n",
"\n",
"word_timestamps = timestamp_dict['word']\n",
"\n",
"for stamp in word_timestamps:\n",
" start = stamp['start_offset'] * time_stride\n",
" end = stamp['end_offset'] * time_stride\n",
" word = stamp['char'] if 'char' in stamp else stamp['word']\n",
"\n",
" print(f\"Time : {start:0.2f} - {end:0.2f} - {word}\")\n",
" display(Audio(signal[int(start * sample_rate) : int(end * sample_rate)], rate=sample_rate))"
],
"metadata": {
"id": "fogttpCTVTEZ"
},
"execution_count": null,
"outputs": []
}
]
}
|