Spaces:
Running
Running
File size: 62,192 Bytes
fb55913 0e7406c 33b4e31 4b0e9d9 fb55913 a9c1bfa 00d10fe 675075b 00d10fe d46b705 00d10fe fb55913 81e551c 6d1f579 fb55913 81e551c fb55913 4b0e9d9 a9c1bfa 0e7406c fb55913 76b58e9 fb55913 76b58e9 4b0e9d9 fb55913 38fa5c5 76b58e9 fb55913 76b58e9 38fa5c5 4b0e9d9 38fa5c5 4b0e9d9 fb55913 4b0e9d9 fb55913 6210a22 c72beff 6210a22 c72beff 6210a22 c72beff 8ba6297 c72beff 96cde07 6210a22 6717298 0f9ddd5 6210a22 8ba6297 6210a22 fb55913 76b58e9 e2c67eb fff497e 6abd7f5 2c6f1eb fff497e 6abd7f5 e2c67eb fff497e e2c67eb fff497e a65930b 6abd7f5 a65930b e2c67eb fff497e 6abd7f5 e2c67eb fff497e e2c67eb 6abd7f5 e2c67eb 6abd7f5 e2c67eb 6abd7f5 e2c67eb fff497e 76b58e9 fff497e 2c6f1eb 76b58e9 dbc189a 76b58e9 950a013 76b58e9 950a013 4752d55 d01d216 4752d55 950a013 800b3ed 950a013 800b3ed d0b6db0 800b3ed 950a013 d0b6db0 02bce6c 6abd7f5 02bce6c 6abd7f5 02bce6c 6abd7f5 02bce6c 76b58e9 d0b6db0 6abd7f5 fb55913 6abd7f5 fb55913 6abd7f5 fb55913 6abd7f5 fb55913 6abd7f5 fb55913 6abd7f5 779a747 6abd7f5 779a747 6abd7f5 779a747 6abd7f5 779a747 6abd7f5 779a747 6abd7f5 779a747 6abd7f5 fb55913 6abd7f5 779a747 6abd7f5 779a747 6abd7f5 779a747 6abd7f5 fb55913 6abd7f5 779a747 6abd7f5 24743e4 6abd7f5 fb55913 6abd7f5 fb55913 6abd7f5 fb55913 6abd7f5 fb55913 6abd7f5 947a29a 6abd7f5 947a29a 6abd7f5 947a29a 6abd7f5 5aee394 6abd7f5 5aee394 6abd7f5 d0b6db0 fb55913 6505f7d 393757e fb55913 393757e fb55913 6505f7d fb55913 0cd1bb4 fb55913 6505f7d fb55913 6505f7d fb55913 6505f7d 6abd7f5 6505f7d 6abd7f5 6505f7d fb55913 6abd7f5 393757e fb55913 6505f7d fb55913 6505f7d fb55913 fff497e fb55913 38fa5c5 fb55913 38fa5c5 bcfd633 38fa5c5 bcfd633 38fa5c5 05d30e4 38fa5c5 f7ef104 fb55913 4b0e9d9 fb55913 4b0e9d9 fb55913 f7ef104 4b0e9d9 fb55913 4b0e9d9 fb55913 8209e58 76b58e9 fb55913 8209e58 fb55913 76b58e9 5dae976 fb55913 38fa5c5 fb55913 0e7406c fb55913 0e7406c fb55913 0e8ac26 fb55913 76b58e9 fb55913 76b58e9 fb55913 6abd7f5 fb55913 779a747 6abd7f5 779a747 b626ce6 fb55913 6abd7f5 fb55913 6abd7f5 fb55913 b626ce6 76b58e9 fb55913 76b58e9 6abd7f5 76b58e9 6210a22 5aee394 6210a22 ced276b 6210a22 ced276b 2ddd203 76b58e9 fb55913 76b58e9 fb55913 76b58e9 fb55913 6abd7f5 33b4e31 76b58e9 393757e 76b58e9 fb55913 393757e b9d73c6 393757e b9d73c6 6abd7f5 b9d73c6 6abd7f5 b9d73c6 6abd7f5 393757e 6abd7f5 76b58e9 4752d55 e28a01e 6abd7f5 fb55913 6abd7f5 fb55913 4752d55 b41d79a fb55913 b41d79a fb55913 6abd7f5 fb55913 6abd7f5 fb55913 6abd7f5 fb55913 76b58e9 fb55913 76b58e9 fb55913 6abd7f5 fb55913 76b58e9 fb55913 |
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 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 |
import os
import io
import subprocess
import numpy as np
from difflib import SequenceMatcher
from datetime import datetime
from typing import List, Tuple, Dict
import asyncio
import base64
import string
import re
import urllib.request
import gzip
import tempfile
# Set cache environment
os.environ['HF_HOME'] = '/tmp/hf'
os.environ['TORCH_HOME'] = '/tmp/torch'
os.environ['TRANSFORMERS_CACHE'] = '/tmp/hf'
os.environ['XDG_CACHE_HOME'] = '/tmp/hf'
os.environ['MPLCONFIGDIR'] = '/tmp/matplotlib' # Fix matplotlib permission issue
os.environ['PULSE_CONFIG_PATH'] = '/tmp/pulse' # Fix PulseAudio errors
os.environ['PULSE_RUNTIME_PATH'] = '/tmp/pulse'
os.makedirs('/tmp/hf', exist_ok=True)
os.makedirs('/tmp/torch', exist_ok=True)
os.makedirs('/tmp/matplotlib', exist_ok=True)
os.makedirs('/tmp/pulse', exist_ok=True)
from fastapi import FastAPI, UploadFile, File, Form
from fastapi.middleware.cors import CORSMiddleware
import torchaudio
import torch
from phonemizer import phonemize
import whisperx # New: WhisperX for precise alignment
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import edge_tts
def log(msg):
print(f"[{datetime.now().strftime('%H:%M:%S')}] {msg}")
app = FastAPI()
app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_credentials=True, allow_methods=["*"], allow_headers=["*"])
def normalize_phoneme_string(s: str) -> str:
"""Normalize phoneme string for comparison - remove spaces and extra chars"""
if not s:
return s
# Convert to lowercase and remove spaces, stress marks, and length markers
normalized = s.lower().strip()
normalized = normalized.replace(' ', '') # Remove spaces between phonemes
normalized = normalized.replace('Λ', '') # Remove length markers
normalized = normalized.replace('Λ', '') # Remove primary stress
normalized = normalized.replace('Λ', '') # Remove secondary stress
normalized = normalized.replace('.', '') # Remove syllable boundaries
# CRITICAL: Explode affricate ligatures into their component parts
# These single-character affricates need to be expanded to match decomposed forms
affricate_ligatures = {
'Κ§': 'tΚ', # Voiceless postalveolar affricate (chip)
'Κ€': 'dΚ', # Voiced postalveolar affricate (jump)
'Κ¦': 'ts', # Voiceless alveolar affricate (German Zeit)
'Κ£': 'dz', # Voiced alveolar affricate (Italian mezzo)
'Κ¨': 'tΙ', # Voiceless alveolo-palatal affricate (Polish Δ)
'Κ₯': 'dΚ', # Voiced alveolo-palatal affricate (Polish dΕΊ)
'Ζ': 'tΙ¬', # Voiceless alveolar lateral affricate (Nahuatl tl)
'κ©': 'tΙ', # Variant for voiceless alveolo-palatal affricate
}
for ligature, expanded in affricate_ligatures.items():
normalized = normalized.replace(ligature, expanded)
# CRITICAL: Normalize ASCII symbols to proper IPA equivalents
# Convert all wav2vec2 ASCII characters to standard IPA
ascii_to_ipa = {
'g': 'Ι‘', # ASCII g β IPA script g (voiced velar stop)
'b': 'b', # ASCII b β IPA b (already correct, but explicit)
'd': 'd', # ASCII d β IPA d (already correct, but explicit)
'f': 'f', # ASCII f β IPA f (already correct, but explicit)
'h': 'h', # ASCII h β IPA h (already correct, but explicit)
'i': 'i', # ASCII i β IPA i (already correct, but explicit)
# Note: Most ASCII phonetic chars are already valid IPA, except 'g'
}
# Normalize variant IPA symbols to consistent forms
# Handle different representations of the same sounds
ipa_variants = {
'Ι': 'Ι', # Open-mid central β r-colored (American English "er")
'Ι': 'Ι', # R-colored schwa β r-colored vowel (both "er" sounds)
'Κ': 'Ι', # Open-mid back β schwa (both unstressed "uh" sounds)
'Γ°': 'ΞΈ', # Voiced th β voiceless th (accent training - treat as equivalent)
'ΙΉ': 'r', # Retroflex approximant β regular r (espeak vs CMUdict difference)
}
for ascii_char, ipa_char in ascii_to_ipa.items():
normalized = normalized.replace(ascii_char, ipa_char)
for variant_char, standard_char in ipa_variants.items():
normalized = normalized.replace(variant_char, standard_char)
return normalized
# Load models once at startup
phoneme_processor = Wav2Vec2Processor.from_pretrained("vitouphy/wav2vec2-xls-r-300m-timit-phoneme")
phoneme_model = Wav2Vec2ForCTC.from_pretrained("vitouphy/wav2vec2-xls-r-300m-timit-phoneme")
# Model inspection complete - wav2vec2 uses ASCII 'g' (token 15), not IPA 'Ι‘'
log("β
Phoneme models loaded - using ASCII/IPA normalization")
# WhisperX models - loaded lazily
whisperx_model = None
whisperx_align_model = None
whisperx_metadata = None
# Simple caches
phoneme_cache = {}
tts_cache = {}
# TTS configuration
TTS_VOICE = "en-US-AriaNeural"
# Phoneme to English letter sounds mapping
PHONEME_TO_ENGLISH = {
# Vowels (monophthongs)
'Ιͺ': 'IH', # "bit"
'Ι': 'EH', # "bed"
'Γ¦': 'AE', # "cat"
'Κ': 'UH', # "but"
'Ι': 'AH', # "father"
'Ι': 'AW', # "law"
'Κ': 'UU', # "book"
'u': 'OO', # "boot"
'i': 'EE', # "beat"
'Ι': 'UH', # "about" (schwa)
'Ι': 'ER', # "bird"
'Ι': 'ER', # "letter"
# Diphthongs
'eΙͺ': 'AY', # "day"
'aΙͺ': 'EYE', # "my"
'ΙΙͺ': 'OY', # "boy"
'aΚ': 'OW', # "now"
'oΚ': 'OH', # "go"
# R-colored vowels
'Ιͺr': 'EER', # "near"
'Ιr': 'AIR', # "care"
'Ιr': 'AR', # "car"
'Ιr': 'OR', # "for"
'Κr': 'OOR', # "tour"
'Ιr': 'ER', # "letter"
'Ι': 'ER', # alternate schwa-r
# Consonants
'p': 'P', # "pat"
'b': 'B', # "bat"
't': 'T', # "tap"
'd': 'D', # "dap"
'k': 'K', # "cat"
'g': 'G', # "gap" (wav2vec2 uses ASCII g)
'Ι‘': 'G', # "gap" (IPA script g - normalize to same)
'f': 'F', # "fat"
'v': 'V', # "vat"
'ΞΈ': 'TH', # "think"
'Γ°': 'TH', # "this"
's': 'S', # "sap"
'z': 'Z', # "zap"
'Κ': 'SH', # "ship"
'Κ': 'ZH', # "measure"
'h': 'H', # "hat"
'm': 'M', # "mat"
'n': 'N', # "nat"
'Ε': 'NG', # "sing"
'l': 'L', # "lap"
'r': 'R', # "rap"
'j': 'Y', # "yes"
'w': 'W', # "wet"
# Affricates
'tΚ': 'CH', # "chip"
'dΚ': 'J', # "jump"
# Common combinations that might appear
' ': '-', # space becomes dash
'Λ': '', # length marker (remove)
'Λ': '', # primary stress (remove)
'Λ': '', # secondary stress (remove)
}
# Phoneme example words - showing the sound in context
PHONEME_EXAMPLES = {
# Vowels (monophthongs)
'Ιͺ': 'bit', # IH sound
'Ι': 'bed', # EH sound
'Γ¦': 'cat', # AE sound
'Κ': 'but', # UH sound (stressed)
'Ι': 'father', # AH sound
'Ι': 'law', # AW sound
'Κ': 'book', # UU sound
'u': 'boot', # OO sound
'i': 'beat', # EE sound
'Ι': 'about', # schwa (unstressed)
'Ι': 'bird', # ER sound (stressed)
'Ι': 'letter', # ER sound (unstressed)
# Diphthongs
'eΙͺ': 'day', # AY sound
'aΙͺ': 'my', # EYE sound
'ΙΙͺ': 'boy', # OY sound
'aΚ': 'now', # OW sound
'oΚ': 'go', # OH sound
# R-colored vowels
'Ιͺr': 'near', # EER sound
'Ιr': 'care', # AIR sound
'Ιr': 'car', # AR sound
'Ιr': 'for', # OR sound
'Κr': 'tour', # OOR sound
'Ιr': 'letter', # ER sound
# Consonants
'p': 'pat', # P sound
'b': 'bat', # B sound
't': 'tap', # T sound
'd': 'dap', # D sound
'k': 'cat', # K sound
'g': 'gap', # G sound (ASCII)
'Ι‘': 'gap', # G sound (IPA)
'f': 'fat', # F sound
'v': 'vat', # V sound
'ΞΈ': 'think', # TH sound (voiceless)
'Γ°': 'this', # TH sound (voiced)
's': 'sap', # S sound
'z': 'zap', # Z sound
'Κ': 'ship', # SH sound
'Κ': 'measure', # ZH sound
'h': 'hat', # H sound
'm': 'mat', # M sound
'n': 'nat', # N sound
'Ε': 'sing', # NG sound
'l': 'lap', # L sound
'r': 'rap', # R sound
'j': 'yes', # Y sound
'w': 'wet', # W sound
# Affricates
'tΚ': 'chip', # CH sound
'dΚ': 'jump', # J sound
}
def clean_word_for_phonemes(word: str) -> str:
"""
Clean word by removing punctuation and extra spaces for phoneme processing.
Keeps only alphabetical characters.
"""
# Remove punctuation and extra whitespace
cleaned = word.strip().translate(str.maketrans('', '', string.punctuation))
cleaned = ''.join(cleaned.split()) # Remove all whitespace
log(f"Word cleaning: '{word}' β '{cleaned}'")
return cleaned
def convert_digits_to_words(text: str) -> str:
"""Convert digits to word form for better phoneme analysis"""
# Dictionary for number conversion
number_words = {
'0': 'zero', '1': 'one', '2': 'two', '3': 'three', '4': 'four',
'5': 'five', '6': 'six', '7': 'seven', '8': 'eight', '9': 'nine',
'10': 'ten', '11': 'eleven', '12': 'twelve', '13': 'thirteen', '14': 'fourteen',
'15': 'fifteen', '16': 'sixteen', '17': 'seventeen', '18': 'eighteen', '19': 'nineteen',
'20': 'twenty', '30': 'thirty', '40': 'forty', '50': 'fifty',
'60': 'sixty', '70': 'seventy', '80': 'eighty', '90': 'ninety',
'100': 'one hundred', '1000': 'one thousand'
}
def convert_number(match):
num_str = match.group()
num = int(num_str)
# Direct lookup for common numbers
if num_str in number_words:
return number_words[num_str]
# Handle numbers 21-99
if 21 <= num <= 99:
tens = (num // 10) * 10
ones = num % 10
if ones == 0:
return number_words[str(tens)]
else:
return number_words[str(tens)] + " " + number_words[str(ones)]
# Handle numbers 101-999 (basic implementation)
if 101 <= num <= 999:
hundreds = num // 100
remainder = num % 100
result = number_words[str(hundreds)] + " hundred"
if remainder > 0:
if remainder < 21:
result += " " + number_words[str(remainder)]
else:
tens = (remainder // 10) * 10
ones = remainder % 10
result += " " + number_words[str(tens)]
if ones > 0:
result += " " + number_words[str(ones)]
return result
# For larger numbers or edge cases, return original
return num_str
# Replace standalone digits/numbers with word equivalents
converted = re.sub(r'\b\d+\b', convert_number, text)
log(f"Number conversion: '{text}' β '{converted}'")
return converted
def load_whisperx_models():
"""Load WhisperX models lazily with English-only configuration"""
global whisperx_model, whisperx_align_model, whisperx_metadata
if whisperx_model is None:
log("Loading WhisperX models for English-only processing...")
# First, try to set environment variable to disable executable stack
os.environ['LD_BIND_NOW'] = '1'
try:
# Try loading with base.en first
whisperx_model = whisperx.load_model("base.en", device="cpu", compute_type="float32", language="en")
log("WhisperX base.en model loaded successfully")
# Load alignment model for English
whisperx_align_model, whisperx_metadata = whisperx.load_align_model(language_code="en", device="cpu")
log("WhisperX English alignment model loaded successfully")
except ImportError as ie:
log(f"Import error loading WhisperX models: {ie}")
# Try to use regular Whisper as fallback
try:
log("Attempting to use standard Whisper instead of WhisperX...")
import whisper
# Load standard whisper model
whisper_model = whisper.load_model("base.en", device="cpu")
# Create a wrapper to make it compatible with WhisperX interface
class WhisperWrapper:
def __init__(self, model):
self.model = model
def transcribe(self, audio, batch_size=16, language="en"):
result = self.model.transcribe(audio, language=language)
# Convert to WhisperX format
return {
"segments": [{
"text": result["text"],
"start": 0.0,
"end": len(audio) / 16000.0, # Approximate based on sample rate
"words": [] # Will need to handle word-level timing differently
}],
"language": language
}
whisperx_model = WhisperWrapper(whisper_model)
log("Using standard Whisper as fallback (limited word-level timing)")
# For alignment, we'll need to handle this differently
whisperx_align_model = None
whisperx_metadata = None
except Exception as whisper_error:
log(f"Standard Whisper fallback failed: {whisper_error}")
# Last resort: Create a minimal mock that at least returns something
class MinimalWhisperMock:
def transcribe(self, audio, batch_size=16, language="en"):
# Return a minimal valid structure
return {
"segments": [{
"text": "[Audio processing unavailable - WhisperX loading failed]",
"start": 0.0,
"end": 1.0,
"words": []
}],
"language": language
}
whisperx_model = MinimalWhisperMock()
whisperx_align_model = None
whisperx_metadata = None
log("WARNING: Using minimal mock - transcription will be limited")
except Exception as e:
log(f"Error loading WhisperX models: {e}")
raise RuntimeError(f"Unable to load speech recognition models: {e}")
def convert_webm_to_wav(bts):
p = subprocess.run(["ffmpeg", "-i", "pipe:0", "-f", "wav", "-ar", "16000", "-ac", "1", "pipe:1"],
input=bts, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
if p.returncode != 0:
raise RuntimeError(p.stderr.decode())
return io.BytesIO(p.stdout)
def calculate_similarity(detected: str, expected: str) -> float:
"""Calculate similarity between detected and expected phonemes"""
detected_norm = normalize_phoneme_string(detected)
expected_norm = normalize_phoneme_string(expected)
return SequenceMatcher(None, detected_norm, expected_norm).ratio()
def detect_word_boundary_overlap(audio_segment: torch.Tensor, sample_rate: int, word: str) -> float:
"""
Analyze first 1/3 of audio segment for: [noise] β [silence] β [noise] pattern
Returns: offset in seconds to skip initial noise, or 0.0 if no pattern found
"""
if audio_segment.shape[-1] == 0:
return 0.0
# Analyze only first 1/3 of segment
first_third_samples = audio_segment.shape[-1] // 3
if first_third_samples < sample_rate * 0.1: # Less than 100ms total
return 0.0
first_third = audio_segment[:, :first_third_samples]
# WORKAROUND: Audio segment appears to be reversed for unknown reason
# This flip corrects the chronological order for proper boundary detection
first_third = torch.flip(first_third, [-1])
# Calculate energy in small windows (50ms chunks)
window_size = int(0.05 * sample_rate) # 50ms windows
if window_size <= 0:
return 0.0
energy_levels = []
for i in range(0, first_third_samples - window_size, window_size):
window = first_third[:, i:i + window_size]
energy = torch.mean(window ** 2).item() # RMS energy
energy_levels.append(energy)
if len(energy_levels) < 3:
return 0.0
# Look for pattern: [high energy] β [low energy] β [high energy]
silence_threshold = np.percentile(energy_levels, 20) # Bottom 20%
noise_threshold = silence_threshold * 3
# Find sustained silence (2+ consecutive low-energy windows)
for i in range(len(energy_levels) - 1):
if (energy_levels[i] < silence_threshold and
energy_levels[i + 1] < silence_threshold):
# Check if there was noise before silence
noise_before = any(e > noise_threshold for e in energy_levels[:i])
# Check if there's noise after silence
noise_after = any(e > noise_threshold for e in energy_levels[i + 2:])
if noise_before and noise_after:
# Found the pattern! Return offset to end of silence
silence_end_sample = (i + 2) * window_size
offset_seconds = silence_end_sample / sample_rate
log(f"π§ Word '{word}': detected boundary overlap, trimming {offset_seconds:.3f}s from start")
return offset_seconds
return 0.0 # No pattern detected
def extract_audio_segment(waveform: torch.Tensor, sample_rate: int,
start_time: float, end_time: float, word: str,
verbose: bool = True) -> torch.Tensor:
"""Extract audio segment for a specific word"""
# Convert to samples
start_sample = int(start_time * sample_rate)
end_sample = int(end_time * sample_rate)
end_sample = min(waveform.shape[-1], end_sample)
if end_sample <= start_sample:
if verbose:
log(f"Invalid segment for '{word}': {start_time:.3f}s-{end_time:.3f}s")
return torch.zeros((1, 1600)) # Return 100ms of silence
segment = waveform[:, start_sample:end_sample]
if verbose:
log(f"Extracted '{word}': {start_time:.3f}s-{end_time:.3f}s ({segment.shape[-1]} samples)")
return segment
def detect_phoneme_from_audio(audio_segment: torch.Tensor, sample_rate: int, word: str) -> str:
"""Detect phoneme from audio segment using phoneme model"""
log(f"π Starting phoneme detection for '{word}'...")
if audio_segment.shape[-1] == 0:
log(f"β οΈ Empty audio segment for '{word}'")
return ""
log(f"π Original audio segment: {audio_segment.shape[-1]} samples")
# Pad or truncate to standard length for model
target_length = 16000 # 1 second
if audio_segment.shape[-1] < target_length:
log(f"π§ Padding audio from {audio_segment.shape[-1]} to {target_length} samples")
audio_segment = torch.nn.functional.pad(audio_segment, (0, target_length - audio_segment.shape[-1]))
elif audio_segment.shape[-1] > target_length:
# Don't truncate long segments - keep full audio for complex words
log(f"β οΈ Audio longer than target ({audio_segment.shape[-1]} > {target_length}), keeping full length")
log(f" This preserves all phonemes for long words like 'sophisticated'")
else:
log(f"β
Audio segment already correct length: {target_length} samples")
log(f"ποΈ Processing through phoneme processor...")
start_time = datetime.now()
# Process through phoneme model
try:
input_values = phoneme_processor(audio_segment.squeeze(), sampling_rate=sample_rate, return_tensors="pt").input_values
processor_time = (datetime.now() - start_time).total_seconds()
log(f"β±οΈ Phoneme processor took: {processor_time:.3f}s")
log(f"π§ Running through phoneme model...")
model_start_time = datetime.now()
with torch.no_grad():
logits = phoneme_model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
detected_phoneme = phoneme_processor.decode(predicted_ids[0])
model_time = (datetime.now() - model_start_time).total_seconds()
log(f"β±οΈ Phoneme model inference took: {model_time:.3f}s")
total_time = (datetime.now() - start_time).total_seconds()
log(f"β±οΈ Total phoneme detection time: {total_time:.3f}s")
except Exception as e:
log(f"β Error in phoneme detection: {e}")
return ""
log(f"π― Phoneme detection for '{word}': '{detected_phoneme}'")
return detected_phoneme
def sliding_window_phoneme_match(detected_phoneme: str, expected_phoneme: str, word: str) -> Tuple[str, float, int, int]:
"""
Find the best matching substring in detected phoneme using sliding window.
For zero scores, intelligently selects which phoneme substring to return.
Returns: (best_match_substring, best_score, start_index, end_index)
"""
detected_norm = normalize_phoneme_string(detected_phoneme)
expected_norm = normalize_phoneme_string(expected_phoneme)
log(f"π Sliding window analysis for '{word}':")
log(f" Expected (norm): '{expected_norm}' (length: {len(expected_norm)})")
log(f" Detected (norm): '{detected_norm}' (length: {len(detected_norm)})")
# If detected is shorter than expected, just compare directly
if len(detected_norm) < len(expected_norm):
score = calculate_similarity(detected_norm, expected_norm)
log(f" Direct comparison (detected < expected): score = {score:.3f}")
return detected_norm, score, 0, len(detected_norm)
# Sliding window: detected is longer than expected
expected_len = len(expected_norm)
best_score = 0
best_match = ""
best_start = 0
best_end = expected_len
log(f" Sliding window search (window size: {expected_len}):")
# Slide through all possible positions
for i in range(len(detected_norm) - expected_len + 1):
substring = detected_norm[i:i + expected_len]
score = calculate_similarity(substring, expected_norm)
log(f" Position {i}: '{substring}' vs '{expected_norm}' = {score:.3f}")
if score > best_score: # Changed back from >= to > to prefer earlier matches
best_score = score
best_match = substring
best_start = i
best_end = i + expected_len
log(f" β
New best match!")
# Exit early on perfect match
if score >= 1.0:
log(f" π― Perfect match found, stopping search")
break
# Handle zero score case - aim for middle substring when possible
if best_score == 0:
log(f" β οΈ Zero score detected, selecting middle substring for audio alignment")
total_detected_len = len(detected_norm)
if total_detected_len == expected_len:
# Same length - use the whole string
best_start = 0
best_end = expected_len
best_match = detected_norm
log(f" π Same length: using full string")
else:
# Longer detected - aim for middle
middle_start = max(0, (total_detected_len - expected_len) // 2)
best_start = middle_start
best_end = middle_start + expected_len
best_match = detected_norm[best_start:best_end]
log(f" π Aiming for middle: position {best_start}-{best_end}")
log(f" π Final selection: '{best_match}' at position {best_start}-{best_end} (score: {best_score:.3f})")
return best_match, best_score, best_start, best_end
def create_word_phoneme_mapping_v2(word: str, expected_phoneme: str) -> Dict[int, str]:
"""
Create mapping from phoneme positions to original word letters.
Simplified version that handles common cases more reliably.
Args:
word: The original word (already cleaned, no punctuation)
expected_phoneme: The expected phoneme string
Returns:
Dictionary mapping phoneme index to word letter(s)
"""
word_lower = word.lower()
phoneme_norm = normalize_phoneme_string(expected_phoneme)
log(f"πΊοΈ Creating mapping for '{word}' β '{phoneme_norm}'")
log(f" Word length: {len(word_lower)}, Phoneme length: {len(phoneme_norm)}")
if not phoneme_norm:
return {}
# Simple cases first
if len(word_lower) == len(phoneme_norm):
# Direct 1:1 mapping
mapping = {i: word[i] for i in range(len(phoneme_norm))} # Preserve original case
log(f" Direct mapping (equal lengths): {mapping}")
return mapping
# For length mismatches, use proportional distribution
mapping = {}
if len(phoneme_norm) > len(word_lower):
# More phonemes than letters (diphthongs, etc.)
# Distribute letters across phonemes without duplication
phonemes_per_letter = len(phoneme_norm) / len(word_lower)
for phoneme_idx in range(len(phoneme_norm)):
# Find which letter this phoneme belongs to
letter_idx = min(int(phoneme_idx / phonemes_per_letter), len(word_lower) - 1)
# Only assign each letter once (to its first phoneme)
start_phoneme_for_letter = int(letter_idx * phonemes_per_letter)
if phoneme_idx == start_phoneme_for_letter:
mapping[phoneme_idx] = word[letter_idx] # Preserve case
else:
mapping[phoneme_idx] = '' # Empty for additional phonemes
else:
# More letters than phonemes (silent letters)
# Distribute letters across available phonemes
letters_per_phoneme = len(word_lower) / len(phoneme_norm)
for phoneme_idx in range(len(phoneme_norm)):
# Calculate range of letters for this phoneme
start_letter = int(phoneme_idx * letters_per_phoneme)
end_letter = int((phoneme_idx + 1) * letters_per_phoneme)
# Collect all letters for this phoneme
letter_group = word[start_letter:end_letter]
mapping[phoneme_idx] = letter_group
log(f" Final mapping: {mapping}")
return mapping
def create_character_level_feedback_v2(word: str, expected_norm: str,
detected_norm: str,
mapping: Dict[int, str]) -> str:
"""
Create character-level feedback with simplified logic.
Args:
word: Original word (for display purposes)
expected_norm: Normalized expected phonemes
detected_norm: Normalized detected phonemes
mapping: Phoneme position to letter mapping
Returns:
HTML string with properly formatted feedback
"""
result = []
log(f"π Character feedback for '{word}':")
log(f" Expected: '{expected_norm}' (len={len(expected_norm)})")
log(f" Detected: '{detected_norm}' (len={len(detected_norm)})")
# Ensure both strings are same length for comparison
max_len = max(len(expected_norm), len(detected_norm))
expected_padded = expected_norm.ljust(max_len)
detected_padded = detected_norm.ljust(max_len)
# Track which word positions have been used
used_positions = set()
for i in range(min(len(expected_norm), max_len)):
expected_char = expected_padded[i] if i < len(expected_padded) else ' '
detected_char = detected_padded[i] if i < len(detected_padded) else ' '
# Get the word letter(s) for this phoneme position
word_letters = mapping.get(i, '')
# Skip empty mappings (extra phonemes in diphthongs)
if not word_letters:
continue
# Check if we've already used these letters
letter_key = (word_letters, i)
if letter_key in used_positions:
continue
used_positions.add(letter_key)
if expected_char == detected_char:
# Correct pronunciation - show original letters
result.append(word_letters)
else:
# Incorrect - create error span with tooltip
expected_english = PHONEME_TO_ENGLISH.get(expected_char, expected_char)
expected_example = PHONEME_EXAMPLES.get(expected_char, '')
detected_english = PHONEME_TO_ENGLISH.get(detected_char, 'silence' if detected_char == ' ' else detected_char)
detected_example = PHONEME_EXAMPLES.get(detected_char, '')
# Build tooltip text
if expected_example and detected_example:
tooltip = f"Expected '{expected_english}' as in '{expected_example}'<br>You said '{detected_english}' as in '{detected_example}'"
elif expected_example:
tooltip = f"Expected '{expected_english}' as in '{expected_example}'<br>You said '{detected_english}'"
else:
tooltip = f"Expected '{expected_english}'<br>You said '{detected_english}'"
# Create error span
error_html = f'<span class="phoneme-error" data-tooltip-html="{tooltip}"><strong><u>{word_letters}</u></strong></span>'
result.append(error_html)
feedback = ''.join(result)
log(f" Final feedback: {feedback}")
return feedback
def format_output_word_v2(word_original: str, word_clean: str,
similarity_score: float, detected_phoneme: str,
expected_phoneme: str, similarity_threshold: float) -> Tuple[str, str]:
"""
Format word output with cleaner logic.
Args:
word_original: Original word with punctuation (for display)
word_clean: Cleaned word (for phoneme processing)
similarity_score: Similarity score between detected and expected
detected_phoneme: Detected phoneme string
expected_phoneme: Expected phoneme string
similarity_threshold: User's threshold for acceptable pronunciation
Returns:
Tuple of (display_text, colored_html)
"""
# Determine color based on score
if similarity_score < similarity_threshold:
color = "red"
needs_feedback = True
elif similarity_score >= similarity_threshold + (1.0 - similarity_threshold) * 0.3:
color = "green"
needs_feedback = False
else:
color = "orange"
needs_feedback = False
score_percentage = int(similarity_score * 100)
if needs_feedback:
# Poor pronunciation - show character-level feedback
# Create phoneme mapping using cleaned word
mapping = create_word_phoneme_mapping_v2(word_clean, expected_phoneme)
# Normalize phonemes for comparison
expected_norm = normalize_phoneme_string(expected_phoneme)
detected_norm = normalize_phoneme_string(detected_phoneme)
# Generate character-level feedback
feedback_html = create_character_level_feedback_v2(
word_clean, expected_norm, detected_norm, mapping
)
# Preserve original punctuation if present
if word_original != word_clean:
# Find trailing punctuation
punct = ''
for i in range(len(word_original) - 1, -1, -1):
if word_original[i] in string.punctuation:
punct = word_original[i] + punct
else:
break
# Find leading punctuation
lead_punct = ''
for char in word_original:
if char in string.punctuation:
lead_punct += char
else:
break
display_text = lead_punct + feedback_html + punct
else:
display_text = feedback_html
# For tooltip, use the cleaned word
tooltip_text = word_clean
else:
# Good pronunciation - show original word
display_text = word_original
tooltip_text = word_original
# Create final colored HTML with embedded data
colored_html = f'<span style="color:{color}" data-score="{score_percentage}" data-word="{word_original}" data-tooltip="{tooltip_text}">{display_text}</span>'
return display_text, colored_html
def trim_audio_segment_by_phoneme_position(audio_segment: torch.Tensor,
detected_phoneme_full: str,
best_start: int, best_end: int,
word: str) -> torch.Tensor:
"""
Trim audio segment based on the position of best matching phoneme substring.
Uses 85% of calculated trim percentages to be less aggressive.
Ensures final segment is never shorter than 0.1 seconds.
Returns the original segment if no trimming is needed.
"""
detected_norm = normalize_phoneme_string(detected_phoneme_full)
total_phoneme_len = len(detected_norm)
if total_phoneme_len == 0 or (best_start == 0 and best_end == total_phoneme_len):
log(f"π΅ No audio trimming needed for '{word}' (using original segment)")
return None # Signal to use original WhisperX timing instead of expanded
# Calculate initial trim percentages
start_trim_pct = best_start / total_phoneme_len
end_trim_pct = (total_phoneme_len - best_end) / total_phoneme_len
# Apply 85% factor to be less aggressive
start_trim_pct_adjusted = start_trim_pct * 0.85
end_trim_pct_adjusted = end_trim_pct * 0.85
# Calculate samples and duration
total_samples = audio_segment.shape[-1]
sample_rate = 16000 # Known sample rate
original_duration = total_samples / sample_rate
# Calculate initial trim amounts
start_trim_samples = int(total_samples * start_trim_pct_adjusted)
end_trim_samples = int(total_samples * end_trim_pct_adjusted)
# Calculate resulting duration
trimmed_samples = total_samples - start_trim_samples - end_trim_samples
trimmed_duration = trimmed_samples / sample_rate
log(f"π΅ Audio trimming for '{word}':")
log(f" Original duration: {original_duration:.3f}s ({total_samples} samples)")
log(f" Phoneme position: {best_start}-{best_end-1} of {total_phoneme_len} chars")
log(f" Initial trim: start={start_trim_pct_adjusted:.1%} ({start_trim_samples} samples), end={end_trim_pct_adjusted:.1%} ({end_trim_samples} samples)")
log(f" Resulting duration: {trimmed_duration:.3f}s")
# MINIMUM DURATION CHECK: Ensure result is at least 0.1 seconds
min_duration = 0.1
min_samples = int(min_duration * sample_rate)
if trimmed_samples < min_samples:
log(f" β οΈ Trimmed duration ({trimmed_duration:.3f}s) below minimum ({min_duration}s)")
# Calculate how much we need to preserve
samples_to_preserve = min_samples
total_trim_needed = total_samples - samples_to_preserve
if total_trim_needed <= 0:
log(f" β οΈ Original segment already at minimum length, no trimming")
return None # Use original WhisperX timing
# Redistribute the trimming proportionally while respecting minimum duration
original_total_trim = start_trim_samples + end_trim_samples
if original_total_trim > 0:
# Scale down both trims proportionally
scale_factor = total_trim_needed / original_total_trim
start_trim_samples = int(start_trim_samples * scale_factor)
end_trim_samples = int(end_trim_samples * scale_factor)
# Ensure we don't exceed total available trim
if start_trim_samples + end_trim_samples > total_trim_needed:
excess = (start_trim_samples + end_trim_samples) - total_trim_needed
# Remove excess from the larger trim
if start_trim_samples > end_trim_samples:
start_trim_samples -= excess
else:
end_trim_samples -= excess
log(f" π§ Adjusted trim: start={start_trim_samples} samples, end={end_trim_samples} samples")
log(f" π§ Scale factor applied: {scale_factor:.3f}")
else:
# Shouldn't happen, but safety check
start_trim_samples = 0
end_trim_samples = 0
# Apply final trimming
trimmed_start = start_trim_samples
trimmed_end = total_samples - end_trim_samples
if trimmed_end <= trimmed_start:
log(f" β οΈ Invalid trim range after adjustment, using original segment")
return None # Signal to use original WhisperX timing
trimmed_segment = audio_segment[:, trimmed_start:trimmed_end]
final_duration = trimmed_segment.shape[-1] / sample_rate
log(f" β
Final result: {trimmed_segment.shape[-1]} samples ({final_duration:.3f}s)")
log(f" β
Trimmed: {start_trim_samples} from start, {end_trim_samples} from end")
return trimmed_segment
def get_expected_phonemes(words: List[str]) -> List[str]:
"""Get expected phonemes using espeak phonemizer"""
cache_key = tuple(words)
if cache_key in phoneme_cache:
log(f"π Using cached phonemes for: {words}")
cached_result = phoneme_cache[cache_key]
log(f" Cached phonemes: {list(zip(words, cached_result))}")
return cached_result
log(f"π€ Getting expected phonemes using phonemizer for: {words}")
try:
# Use espeak phonemizer to get IPA phonemes
phonemes = phonemize(words, language='en-us', backend='espeak', strip=True)
# Cache the results
phoneme_cache[cache_key] = phonemes
# Log the phoneme results
log(f"β
Phonemizer results:")
for word, phoneme in zip(words, phonemes):
log(f" '{word}' β '{phoneme}'")
return phonemes
except Exception as e:
log(f"β Error in phonemizer: {e}")
log(f" Returning empty phonemes for all words")
# Return empty strings as fallback
empty_results = [""] * len(words)
phoneme_cache[cache_key] = empty_results
return empty_results
async def generate_tts_audio(word: str) -> str:
"""Generate TTS audio for a word with silence padding"""
if word in tts_cache:
return tts_cache[word]
try:
communicate = edge_tts.Communicate(word, TTS_VOICE)
audio_data = b""
async for chunk in communicate.stream():
if chunk["type"] == "audio":
audio_data += chunk["data"]
if audio_data:
# Add silence padding to TTS audio as well
# First decode the MP3 to get raw audio
import tempfile
with tempfile.NamedTemporaryFile(suffix='.mp3', delete=False) as tmp_mp3:
tmp_mp3.write(audio_data)
tmp_mp3_path = tmp_mp3.name
try:
# Load the TTS audio
tts_waveform, tts_sample_rate = torchaudio.load(tmp_mp3_path)
# Resample if needed to match our standard rate
if tts_sample_rate != 16000:
tts_waveform = torchaudio.transforms.Resample(tts_sample_rate, 16000)(tts_waveform)
tts_sample_rate = 16000
# Add 0.25s silence padding on each end
padding_samples = int(0.25 * tts_sample_rate)
silence_shape = list(tts_waveform.shape)
silence_shape[-1] = padding_samples
silence_padding = torch.zeros(silence_shape)
# Concatenate: silence + audio + silence
padded_waveform = torch.cat([silence_padding, tts_waveform, silence_padding], dim=-1)
# Convert back to base64
buffer = io.BytesIO()
torchaudio.save(buffer, padded_waveform, tts_sample_rate, format="wav")
buffer.seek(0)
audio_b64 = base64.b64encode(buffer.read()).decode('utf-8')
tts_cache[word] = audio_b64
log(f"π TTS for '{word}': Added 0.25s silence padding on each end")
return audio_b64
finally:
# Clean up temp file
if os.path.exists(tmp_mp3_path):
os.remove(tmp_mp3_path)
except Exception as e:
log(f"TTS failed for '{word}': {e}")
return ""
def audio_to_base64(audio_segment: torch.Tensor, sample_rate: int, add_padding: bool = True) -> str:
"""
Convert audio tensor to base64 string.
Args:
audio_segment: The audio tensor to convert
sample_rate: Sample rate of the audio
add_padding: If True, adds 0.25s of silence on each end to prevent audio processor lag
Returns:
Base64 encoded audio string
"""
try:
if add_padding:
# Add 0.25 seconds of silence on each end
padding_samples = int(0.25 * sample_rate) # 0.25 seconds worth of samples
# Create silence padding (zeros with same shape as audio segment)
silence_shape = list(audio_segment.shape)
silence_shape[-1] = padding_samples
silence_padding = torch.zeros(silence_shape)
# Concatenate: silence + audio + silence
padded_segment = torch.cat([silence_padding, audio_segment, silence_padding], dim=-1)
log(f"π Added silence padding: {padding_samples} samples (0.25s) on each end")
log(f" Original: {audio_segment.shape[-1]} samples β Padded: {padded_segment.shape[-1]} samples")
audio_segment = padded_segment
buffer = io.BytesIO()
torchaudio.save(buffer, audio_segment, sample_rate, format="wav")
buffer.seek(0)
return base64.b64encode(buffer.read()).decode('utf-8')
except Exception as e:
log(f"Audio conversion failed: {e}")
return ""
@app.post("/api/transcribe")
async def transcribe(audio: UploadFile = File(...), similarity_threshold: float = Form(0.4)):
log("=== STARTING WHISPERX ENGLISH-ONLY PHONEME ANALYSIS ===")
# Use similarity threshold from frontend (default 0.4)
similarity = max(0.0, min(1.0, similarity_threshold)) # Clamp between 0 and 1
log(f"Using similarity threshold: {similarity:.2f}")
try:
# Load WhisperX models if needed
load_whisperx_models()
# 1. Convert and load audio
data = await audio.read()
wav_io = convert_webm_to_wav(data)
# Save to temporary file for WhisperX
temp_audio_path = "/tmp/temp_audio.wav"
with open(temp_audio_path, "wb") as f:
f.write(wav_io.getvalue())
# Load audio with WhisperX
audio_data = whisperx.load_audio(temp_audio_path)
log(f"Audio loaded for WhisperX: {len(audio_data)} samples")
# 2. Get transcription with WhisperX - EXPLICITLY SET TO ENGLISH
result = whisperx_model.transcribe(audio_data, batch_size=16, language="en")
# 3. Get precise word alignments with WhisperX (if alignment model available)
if whisperx_align_model is not None:
aligned_result = whisperx.align(result["segments"], whisperx_align_model, whisperx_metadata, audio_data, device="cpu")
else:
log("WARNING: Alignment model not available, using basic word splitting")
# Fallback: split text into words with approximate timing
aligned_result = {"segments": []}
for segment in result["segments"]:
text = segment.get("text", "").strip()
if not text:
continue
words = text.split()
duration = segment["end"] - segment["start"]
time_per_word = duration / len(words) if words else 0
word_list = []
for i, word in enumerate(words):
word_start = segment["start"] + (i * time_per_word)
word_end = segment["start"] + ((i + 1) * time_per_word)
word_list.append({
"word": word,
"start": word_start,
"end": word_end,
"score": 0.9 # Default confidence
})
aligned_result["segments"].append({
"text": text,
"start": segment["start"],
"end": segment["end"],
"words": word_list
})
# Extract word-level data from WhisperX results
words = []
word_texts = [] # Original with punctuation for display
word_texts_clean = [] # Cleaned for phoneme processing
word_timings = []
for segment in aligned_result["segments"]:
if "words" in segment:
for word_info in segment["words"]:
if "start" in word_info and "end" in word_info and word_info["word"]:
original_word = word_info["word"].strip()
# Convert digits to words for better phoneme analysis
word_converted = convert_digits_to_words(original_word)
cleaned_word = clean_word_for_phonemes(word_converted)
# Only process words that have alphabetical content after cleaning
if cleaned_word:
words.append(word_info)
word_texts.append(word_converted) # Use converted form for display
word_texts_clean.append(cleaned_word) # Clean for processing
word_timings.append((word_info["start"], word_info["end"]))
if not words:
return {"resolved": "", "resolved_colored": "", "audio_data": []}
log(f"Found {len(words)} words with precise WhisperX timings")
# Log WhisperX timings
log("=== WHISPERX PRECISE TIMINGS ===")
for i, (word_original, word_clean, (start, end)) in enumerate(zip(word_texts, word_texts_clean, word_timings)):
gap = ""
if i > 0:
prev_end = word_timings[i-1][1]
gap_duration = start - prev_end
gap = f" | gap: {gap_duration:.3f}s"
log(f"Word {i}: '{word_original}' (clean: '{word_clean}') at {start:.3f}s-{end:.3f}s{gap}")
# 4. Get expected phonemes using CLEANED words
expected_phonemes = get_expected_phonemes(word_texts_clean)
# 5. Load audio as tensor for phoneme analysis
waveform, sample_rate = torchaudio.load(temp_audio_path)
if sample_rate != 16000:
waveform = torchaudio.transforms.Resample(sample_rate, 16000)(waveform)
sample_rate = 16000
# 6. Process each word using expanded timing with sliding window matching
results = []
audio_data_list = []
# Generate TTS for all words concurrently (using CLEANED words)
log("Generating TTS audio...")
tts_tasks = [generate_tts_audio(word_clean) for word_clean in word_texts_clean]
tts_results = await asyncio.gather(*tts_tasks)
log("\n=== PROCESSING WORDS WITH EXPANDED TIMING + SLIDING WINDOW ===")
for i, (word_info, word_original, word_clean, (start_time, end_time)) in enumerate(zip(words, word_texts, word_texts_clean, word_timings)):
expected_phoneme = expected_phonemes[i] if i < len(expected_phonemes) else ""
log(f"\n--- Processing word {i}: '{word_original}' (clean: '{word_clean}') ---")
log(f"π WhisperX timing: {start_time:.3f}s - {end_time:.3f}s (duration: {end_time - start_time:.3f}s)")
log(f"π― Expected phoneme: '{expected_phoneme}'")
# DEBUGGING: Special attention to problematic words
if word_clean.lower() in ['go', 'no', 'so', 'to', 'do'] or len(expected_phoneme) > len(word_clean):
log(f"β οΈ SPECIAL CASE: Word '{word_clean}' has {len(word_clean)} letters but {len(expected_phoneme)} phonemes")
log(f" This may be a diphthong case requiring special handling")
# For very short words, expand the WhisperX timing itself before processing
original_duration = end_time - start_time
if original_duration < 0.1:
log(f"π Ultra-short word detected ({original_duration:.3f}s), expanding WhisperX timing")
audio_duration = waveform.shape[-1] / sample_rate
# Expand WhisperX boundaries by Β±0.05s
start_time = max(0, start_time - 0.05)
end_time = min(audio_duration, end_time + 0.05)
log(f" Expanded WhisperX timing: {start_time:.3f}s - {end_time:.3f}s (new duration: {end_time - start_time:.3f}s)")
# Show gaps between words
if i > 0:
prev_end = word_timings[i-1][1]
gap = start_time - prev_end
if gap > 0:
log(f"βΈοΈ Gap from previous word: {gap:.3f}s")
elif gap < 0:
log(f"β οΈ OVERLAP with previous word: {gap:.3f}s")
else:
log(f"π No gap (continuous)")
# Calculate expanded timing (Β±0.125s with boundary protection)
expansion_seconds = 0.125
audio_duration = waveform.shape[-1] / sample_rate
expanded_start = max(0, start_time - expansion_seconds)
expanded_end = min(audio_duration, end_time + expansion_seconds)
log(f"π Timing expansion: {start_time:.3f}s-{end_time:.3f}s β {expanded_start:.3f}s-{expanded_end:.3f}s")
# Extract expanded audio segment
expanded_audio_segment = extract_audio_segment(waveform, sample_rate, expanded_start, expanded_end, word_clean, verbose=True)
# Check for word boundary overlap and trim if needed
log(f"π Checking word boundary overlap for '{word_clean}'...")
boundary_offset = detect_word_boundary_overlap(expanded_audio_segment, sample_rate, word_clean)
if boundary_offset > 0:
log(f"π§ Detected word overlap, trimming {boundary_offset:.3f}s from start")
trim_samples = int(boundary_offset * sample_rate)
expanded_audio_segment = expanded_audio_segment[:, trim_samples:]
# Update expanded_start for accurate timing logs
expanded_start += boundary_offset
log(f" Updated expanded start: {expanded_start:.3f}s")
# ALSO apply the boundary offset to WhisperX timing
original_start_time = start_time
start_time = max(0, start_time + boundary_offset)
end_time = max(start_time + 0.01, end_time) # Ensure minimum 10ms duration
log(f" Updated WhisperX timing: {original_start_time:.3f}s β {start_time:.3f}s (shifted +{boundary_offset:.3f}s)")
# Also extract WhisperX original timing for comparison (now using updated start_time)
whisperx_audio_segment = extract_audio_segment(waveform, sample_rate, start_time, end_time, word_clean, verbose=False)
# Detect phoneme from expanded audio segment
detected_phoneme_raw = detect_phoneme_from_audio(expanded_audio_segment, sample_rate, word_clean)
# Get expected phoneme and normalize both
detected_phoneme_norm = normalize_phoneme_string(detected_phoneme_raw)
expected_phoneme_norm = normalize_phoneme_string(expected_phoneme)
log(f"π Raw detected phoneme (expanded): '{detected_phoneme_raw}'")
log(f"π§Ή Normalized detected: '{detected_phoneme_norm}'")
log(f"π§Ή Normalized expected: '{expected_phoneme_norm}'")
# Find best matching substring using sliding window
best_match_phoneme, similarity_score, match_start, match_end = sliding_window_phoneme_match(
detected_phoneme_raw, expected_phoneme, word_clean
)
log(f"π Final similarity score: {similarity_score:.3f}")
# Trim audio segment based on best phoneme match position
trimmed_audio_segment = trim_audio_segment_by_phoneme_position(
expanded_audio_segment, detected_phoneme_raw, match_start, match_end, word_clean
)
# Use original WhisperX timing if no trimming was needed, otherwise use trimmed
if trimmed_audio_segment is None:
final_audio_segment = whisperx_audio_segment
log(f"π΅ Using original WhisperX timing (no trimming needed)")
log(f" Final segment: WhisperX original ({whisperx_audio_segment.shape[-1]} samples, {whisperx_audio_segment.shape[-1]/sample_rate:.3f}s)")
log(f" Segment timing: {start_time:.3f}s - {end_time:.3f}s")
else:
final_audio_segment = trimmed_audio_segment
log(f"π΅ Using trimmed segment from expanded timing")
log(f" Final segment: Processed ({trimmed_audio_segment.shape[-1]} samples, {trimmed_audio_segment.shape[-1]/sample_rate:.3f}s)")
# Calculate the actual timing of the processed segment
final_duration = trimmed_audio_segment.shape[-1] / sample_rate
expanded_duration = expanded_end - expanded_start
# Calculate trim amounts based on phoneme positions
detected_phoneme_norm = normalize_phoneme_string(detected_phoneme_raw)
total_phoneme_len = len(detected_phoneme_norm)
if total_phoneme_len > 0:
start_trim_pct = match_start / total_phoneme_len * 0.85 # Apply 85% factor
end_trim_pct = (total_phoneme_len - match_end) / total_phoneme_len * 0.85
time_trimmed_from_start = expanded_duration * start_trim_pct
time_trimmed_from_end = expanded_duration * end_trim_pct
final_start_time = expanded_start + time_trimmed_from_start
final_end_time = expanded_end - time_trimmed_from_end
log(f" Segment timing: {final_start_time:.3f}s - {final_end_time:.3f}s")
else:
log(f" Segment timing: {expanded_start:.3f}s - {expanded_end:.3f}s (no phoneme-based calculation)")
log(f"π Audio segments returned to user:")
log(f" 1οΈβ£ Expected (TTS): Generated speech")
log(f" 2οΈβ£ User audio: {'WhisperX original' if trimmed_audio_segment is None else 'Processed/trimmed'} ({final_audio_segment.shape[-1]} samples)")
log(f" 3οΈβ£ WhisperX raw: Original timing ({whisperx_audio_segment.shape[-1]} samples)")
if trimmed_audio_segment is not None and final_audio_segment.shape[-1] != whisperx_audio_segment.shape[-1]:
sample_diff = final_audio_segment.shape[-1] - whisperx_audio_segment.shape[-1]
time_diff = sample_diff / sample_rate
log(f" π Segment difference: {sample_diff:+d} samples ({time_diff:+.3f}s) processed vs WhisperX")
log(f"π Final similarity score: {similarity_score:.3f}")
log(f"π¨ Final audio segment samples: {final_audio_segment.shape[-1]} (duration: {final_audio_segment.shape[-1]/sample_rate:.3f}s)")
log(f"π€ WhisperX original segment samples: {whisperx_audio_segment.shape[-1]} (duration: {whisperx_audio_segment.shape[-1]/sample_rate:.3f}s)")
log(f"β° Timing comparison:")
log(f" WhisperX original: {start_time:.3f}s - {end_time:.3f}s (duration: {end_time - start_time:.3f}s)")
log(f" Expanded timing: {expanded_start:.3f}s - {expanded_end:.3f}s (duration: {expanded_end - expanded_start:.3f}s)")
# Store results - now with both original and clean versions
results.append({
'word_original': word_original, # Original with punctuation for display
'word_clean': word_clean, # Cleaned version for phoneme processing
'detected_phoneme': best_match_phoneme, # Use best matching substring
'expected_phoneme': expected_phoneme,
'similarity_score': float(similarity_score),
'start_time': float(start_time),
'end_time': float(end_time),
'whisperx_confidence': float(word_info.get('score', 1.0))
})
# Prepare audio data with all three segments (use ORIGINAL word for display)
# All three audio segments will have 0.25s silence padding added automatically
user_audio_b64 = audio_to_base64(final_audio_segment, sample_rate) # Padded
whisperx_audio_b64 = audio_to_base64(whisperx_audio_segment, sample_rate) # Padded
expected_audio_b64 = tts_results[i] # Already padded in generate_tts_audio
audio_data_list.append({
"word": word_original, # Original with punctuation for display
"expected_audio": expected_audio_b64, # TTS with padding
"user_audio": user_audio_b64, # User's pronunciation with padding
"whisperx_audio": whisperx_audio_b64, # WhisperX original with padding
"start_time": float(start_time),
"end_time": float(end_time),
"similarity_score": float(similarity_score),
"detected_phoneme": best_match_phoneme, # Use best matching substring
"expected_phoneme": expected_phoneme,
"whisperx_confidence": float(word_info.get('score', 1.0))
})
# 7. Format output using the refactored v2 functions
resolved_output = []
resolved_colored = []
for result in results:
output_text, colored_text = format_output_word_v2(
result['word_original'], # Pass both versions
result['word_clean'],
result['similarity_score'],
result['detected_phoneme'],
result['expected_phoneme'],
similarity
)
resolved_output.append(output_text)
resolved_colored.append(colored_text)
# Clean up temporary file
os.remove(temp_audio_path)
log("=== WHISPERX ENGLISH-ONLY PHONEME ANALYSIS COMPLETE ===")
return {
"resolved": " ".join(resolved_output),
"resolved_colored": " ".join(resolved_colored),
"audio_data": audio_data_list,
"debug_info": {
"total_words": len(words),
"similarity_threshold": similarity,
"alignment_method": "WhisperX English-only + Sliding Window",
"results_summary": [
{
"word": r['word_original'],
"score": float(r['similarity_score']),
"detected": r['detected_phoneme'],
"expected": r['expected_phoneme'],
"whisperx_confidence": r['whisperx_confidence']
}
for r in results
]
}
}
except Exception as e:
log(f"ERROR in transcribe: {str(e)}")
import traceback
log(f"Traceback: {traceback.format_exc()}")
return {
"resolved": "Error occurred",
"resolved_colored": "Error occurred",
"audio_data": [],
"debug_info": {"error": str(e)}
}
@app.get("/")
def root():
return "Clean Fonetik with WhisperX English-only + Character-Level Feedback running"
@app.post("/api/clear-cache")
def clear_cache():
global phoneme_cache, tts_cache
phoneme_cache.clear()
tts_cache.clear()
return {"message": "Cache cleared"} |