File size: 30,770 Bytes
be3e6cf b9b7670 f75eacf c767c25 b9b7670 f75eacf be3e6cf f75eacf 8489329 f75eacf 8489329 60cff39 8489329 5d519e9 8489329 5d519e9 8489329 be3e6cf 8489329 b9b7670 be3e6cf f75eacf 8489329 f75eacf 8489329 f75eacf be3e6cf f75eacf 8489329 be3e6cf 8489329 f75eacf 8489329 f75eacf 8489329 be3e6cf f75eacf 8489329 f75eacf 5d519e9 f75eacf 5d519e9 f75eacf be3e6cf f75eacf be3e6cf f75eacf 5d519e9 f75eacf b9b7670 f75eacf be3e6cf f75eacf b9b7670 f75eacf 5d519e9 f75eacf 5d519e9 f75eacf 5d519e9 be3e6cf 5d519e9 f75eacf 5d519e9 f75eacf 5d519e9 f75eacf 5d519e9 f75eacf b9b7670 f75eacf b9b7670 f75eacf c767c25 f75eacf 5d519e9 f75eacf 5d519e9 f75eacf 5d519e9 f75eacf 5d519e9 3d93605 5d519e9 3d93605 5d519e9 f75eacf 4fffe95 5d519e9 4fffe95 5d519e9 f75eacf be3e6cf 5d519e9 f75eacf 4fffe95 f75eacf 5d519e9 f75eacf 4fffe95 f75eacf 4fffe95 be3e6cf f75eacf 5d519e9 f75eacf 5d519e9 f75eacf c767c25 5d519e9 f75eacf 5d519e9 be3e6cf 4fffe95 f75eacf 4fffe95 f75eacf 5d519e9 f75eacf 5d519e9 f75eacf 5d519e9 4fffe95 f75eacf 5d519e9 f75eacf 5d519e9 4fffe95 5d519e9 4fffe95 5d519e9 4fffe95 5d519e9 4fffe95 f75eacf 5d519e9 f75eacf 5d519e9 4fffe95 5d519e9 be3e6cf 5d519e9 8489329 c767c25 8489329 f75eacf be3e6cf 8489329 be3e6cf 8489329 be3e6cf 5d519e9 be3e6cf 8489329 be3e6cf 5d519e9 be3e6cf 5d519e9 4fffe95 5d519e9 be3e6cf 8489329 be3e6cf 5d519e9 be3e6cf 4fffe95 5d519e9 be3e6cf b9b7670 5d519e9 be3e6cf 8489329 be3e6cf 5d519e9 be3e6cf 8489329 5d519e9 be3e6cf 8489329 be3e6cf f75eacf be3e6cf c767c25 f75eacf c767c25 f75eacf be3e6cf c767c25 f75eacf 8489329 f75eacf 8489329 f75eacf c767c25 f75eacf c767c25 f75eacf be3e6cf c767c25 f75eacf be3e6cf f75eacf c767c25 |
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 |
import os, re, json, math, tempfile, traceback
import numpy as np
import pandas as pd
import torch
import textdistance
import gradio as gr
from faster_whisper import WhisperModel
from sentence_transformers import SentenceTransformer, util
from transformers import AutoTokenizer, AutoModel
import soundfile as sf
# =========================
# Global config (forced per your request)
# =========================
FORCE_WHISPER_NAME = "large-v3"
FORCE_COMPUTE_TYPE = "int8"
FORCE_USE_MARBERT = True
# ======= Budget Config =======
# "auto": ูุนุชู
ุฏ ุนูู ุงูุญุงุฑุณ ุงูุนุงูู
ู (SBERT/ROUGE/WER)
# "fixed": ุนุฏุฏ ุซุงุจุช ู
ู ุงูุงุณุชุจุฏุงูุงุช (0 ูุนูู ุนุฏู
ุงุณุชุจุฏุงู ู
ุทูููุง)
# "ratio": ูุณุจุฉ ู
ู ุทูู ุงููุต ุงูู
ูุทูู
# "off": ุจุฏูู ุณูู (ุณููู ูุฏูู
)
FORCE_BUDGET_MODE = "auto" # "auto" | "fixed" | "ratio" | "off"
FIXED_BUDGET_TOKENS = 0
BUDGET_RATIO = 0.15
# =============================
# ุฎูุงุฑุงุช ุชูุฑูุบ ุซุงุจุชุฉ ูุชูููู ุงููุฑููุงุช
ASR_OPTS = dict(
word_timestamps=True,
vad_filter=True,
vad_parameters={"min_silence_duration_ms": 200},
beam_size=5,
best_of=5,
temperature=0.0,
)
# =========================
# Device
# =========================
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"[INIT] DEVICE={DEVICE}", flush=True)
# =========================
# Lazy models
# =========================
_SBERT = None
_MARBERT_TOK = None
_MARBERT = None
_WHISPER = None
def load_models(
sbert_name="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
marbert_name="UBC-NLP/MARBERT",
whisper_name=FORCE_WHISPER_NAME,
whisper_compute=FORCE_COMPUTE_TYPE,
use_marbert=FORCE_USE_MARBERT
):
"""Load models once; forced config respected even on CPU."""
global _SBERT, _MARBERT_TOK, _MARBERT, _WHISPER
if _SBERT is None:
_SBERT = SentenceTransformer(sbert_name, device=("cuda" if DEVICE=="cuda" else "cpu"))
print(f"[LOAD] SBERT: {sbert_name}", flush=True)
if _MARBERT is None and use_marbert:
_MARBERT_TOK = AutoTokenizer.from_pretrained(marbert_name)
_MARBERT = AutoModel.from_pretrained(marbert_name).to(("cuda" if DEVICE=="cuda" else "cpu"))
_MARBERT.eval()
print(f"[LOAD] MARBERT: {marbert_name} (device={DEVICE})", flush=True)
if _WHISPER is None:
_WHISPER = WhisperModel(whisper_name, device=("cuda" if DEVICE=="cuda" else "cpu"),
compute_type=whisper_compute)
print(f"[LOAD] Whisper: {whisper_name} (compute={whisper_compute})", flush=True)
# =========================
# Normalization / Tokenization / Alignment
# =========================
def normalize_ar_orth(text: str) -> str:
# ุชุทุจูุน ุนุงู
ููู
ุญุงุฐุงุฉ
text = re.sub(r"[ููููููููู]", "", text)
text = re.sub(r"[โโ\"',:ุุ.!()\[\]{}ุ\-โโ_]", " ", text)
text = re.sub(r"\s+", " ", text).strip()
return text
def _normalize_for_models(s: str) -> str:
# ุชุทุจูุน ุฎุงุต ูู
ุฏุฎูุงุช SBERT/MARBERT
s = re.sub(r"[ููููููููู]", "", s)
s = re.sub(r"[โโ\"',:ุุ.!()\[\]{}ุ\-โโ_]", " ", s)
s = re.sub(r"\s+", " ", s).strip()
return s
def simple_tokenize(text: str):
t = normalize_ar_orth(text)
try:
import nltk
try:
nltk.data.find('tokenizers/punkt')
except LookupError:
nltk.download('punkt', quiet=True)
return nltk.word_tokenize(t)
except Exception:
return t.split()
def align_texts(ref_tokens, hyp_tokens):
import difflib
sm = difflib.SequenceMatcher(None, ref_tokens, hyp_tokens)
aligned = []
for tag, i1, i2, j1, j2 in sm.get_opcodes():
aligned.append({
'type': tag,
'ref': ref_tokens[i1:i2],
'hyp': hyp_tokens[j1:j2],
'ref_idx': (i1, i2),
'hyp_idx': (j1, j2)
})
return aligned
# =========================
# Phonetic / Levenshtein
# =========================
def arabic_soundex(word):
w = normalize_ar_orth(word)
groups = {
'b': 'ุจู', 'j': 'ุฌุดุต', 'd': 'ุฏุถ', 't': 'ุทุช', 'q': 'ูุบ', 'k': 'ูุฎ',
's': 'ุณุตุฒ', 'z': 'ุซุฐุธ', 'h': 'ุญ', 'a': 'ุน', 'm': 'ู
', 'n': 'ู',
'l': 'ู', 'r': 'ุฑ', 'w': 'ู', 'y': 'ู'
}
code = []
for ch in w:
for rep, chars in groups.items():
if ch in chars:
code.append(rep); break
return "".join(code)
def phonetic_similarity(w1, w2):
if not w1 or not w2: return False
return arabic_soundex(w1) == arabic_soundex(w2)
def is_levenshtein_1(w1, w2):
if not w1 or not w2: return False
return textdistance.levenshtein(w1, w2) == 1
# =========================
# Numbers
# =========================
AR_DIGITS = str.maketrans("ู ูกูขูฃูคูฅูฆูงูจูฉ", "0123456789")
UNITS = {"ุตูุฑ":0,"ูุงุญุฏ":1,"ูุงุญุฏุฉ":1,"ุงุซูุงู":2,"ุงุซููู":2,"ุงุซูุชุงู":2,"ุงุซูุชูู":2,
"ุซูุงุซ":3,"ุซูุงุซู":3,"ุซูุงุซุฉ":3,"ุงุฑุจุน":4,"ุงุฑุจุนู":4,"ุฃุฑุจุน":4,"ุฃุฑุจุนู":4,
"ุฎู
ุณ":5,"ุฎู
ุณู":5,"ุณุช":6,"ุณุชู":6,"ุณุจุน":7,"ุณุจุนู":7,"ุซู
ุงู":8,"ุซู
ุงูู":8,"ุซู
ุงููู":8,
"ุชุณุน":9,"ุชุณุนู":9}
TENS = {"ุนุดุฑ":10,"ุนุดุฑุฉ":10,"ุนุดุฑู":10,"ุนุดุฑูู":20,"ุนุดุฑูู":20,"ุซูุงุซูู":30,"ุซูุงุซูู":30,
"ุงุฑุจุนูู":40,"ุฃุฑุจุนูู":40,"ุงุฑุจุนูู":40,"ุฎู
ุณูู":50,"ุณุชูู":60,"ุณุจุนูู":70,"ุซู
ุงููู":80,"ุชุณุนูู":90}
HUND = {"ู
ุฆู":100,"ู
ุฆุฉ":100,"ู
ุงุฆู":100}
SCALE = {"ุงูู":1000,"ุฃูู":1000,"ุขูุงู":1000,"ู
ูููู":10**6,"ู
ููุงุฑ":10**9}
def normalize_digits(s: str) -> str:
return s.translate(AR_DIGITS)
def words_to_number(tokens):
total = 0; current = 0
for w in tokens:
if w in UNITS: current += UNITS[w]
elif w in TENS: current += TENS[w]
elif w in HUND: current += HUND[w]
elif w in SCALE:
current = max(1, current) * SCALE[w]
total += current; current = 0
elif w == "ู":
continue
else:
total += current; current = 0
total += current
return total if total != 0 else None
def to_numeric_value(token: str):
if not token: return None
t = normalize_ar_orth(token)
d = normalize_digits(t)
if re.fullmatch(r"\d+", d):
return int(d)
toks = t.split()
return words_to_number(toks)
# =========================
# Semantic similarities (MARBERT fixed)
# =========================
def _mean_pool(last_hidden_state, attention_mask):
mask = attention_mask.unsqueeze(-1).expand(last_hidden_state.size()).float()
summed = (last_hidden_state * mask).sum(dim=1)
counts = mask.sum(dim=1).clamp(min=1e-9)
return summed / counts
def marbert_cls_similarity(a: str, b: str) -> float:
"""Return 0 when [UNK] dominates; use mean pooling instead of CLS only."""
if not a or not b or _MARBERT is None:
return 0.0
a_n = _normalize_for_models(a)
b_n = _normalize_for_models(b)
# UNK ratio check
ids_a = _MARBERT_TOK(a_n, add_special_tokens=False).input_ids
ids_b = _MARBERT_TOK(b_n, add_special_tokens=False).input_ids
unk_id = _MARBERT_TOK.unk_token_id
if len(ids_a) == 0 or len(ids_b) == 0:
return 0.0
unk_ratio_a = (ids_a.count(unk_id) / len(ids_a)) if unk_id is not None else 0.0
unk_ratio_b = (ids_b.count(unk_id) / len(ids_b)) if unk_id is not None else 0.0
if max(unk_ratio_a, unk_ratio_b) > 0.5:
# too many unknowns โ ignore MARBERT
return 0.0
with torch.no_grad():
ta = _MARBERT_TOK(a_n, return_tensors='pt', truncation=True, padding=True).to(("cuda" if DEVICE=="cuda" else "cpu"))
tb = _MARBERT_TOK(b_n, return_tensors='pt', truncation=True, padding=True).to(("cuda" if DEVICE=="cuda" else "cpu"))
ea = _mean_pool(_MARBERT(**ta).last_hidden_state, ta["attention_mask"])
eb = _mean_pool(_MARBERT(**tb).last_hidden_state, tb["attention_mask"])
sim = util.cos_sim(ea, eb).item() # -1..1
return (sim + 1) / 2 # 0..1
def multi_bert_similarity(a: str, b: str):
if not a or not b:
return {"sbert":0.0, "marbert":0.0, "max":0.0, "avg":0.0, "note":"empty"}
a_n = _normalize_for_models(a); b_n = _normalize_for_models(b)
sbert_sim = float(util.pytorch_cos_sim(
_SBERT.encode(a_n, convert_to_tensor=True),
_SBERT.encode(b_n, convert_to_tensor=True)
))
marbert_sim = marbert_cls_similarity(a_n, b_n)
note = None
if abs(sbert_sim - marbert_sim) > 0.35:
note = "models_disagree"
vals = [sbert_sim, marbert_sim]
return {"sbert": sbert_sim, "marbert": marbert_sim,
"max": max(vals), "avg": sum(vals)/len(vals), "note": note}
# =========================
# Faster-Whisper helpers
# =========================
def clean_ar_token(t: str) -> str:
t = t.strip()
t = re.sub(r'^[^\w\u0600-\u06FF]+|[^\w\u0600-\u06FF]+$', '', t)
t = normalize_ar_orth(t)
return t
def extract_word_conf_table(segments):
rows = []
for seg in segments:
for w in (seg.words or []):
rows.append({
"seg_start": float(seg.start),
"seg_end": float(seg.end),
"word_start": float(w.start),
"word_end": float(w.end),
"word": clean_ar_token(w.word),
"prob": float(w.probability),
})
return pd.DataFrame(rows)
def build_asr_token_conf(df_words: pd.DataFrame, hyp_tokens: list):
toks_probs, toks_durs = [], []
for _, row in df_words.iterrows():
prob = row["prob"]
dur = (row["word_end"] - row["word_start"]) * 1000.0
toks_probs.append(prob)
toks_durs.append(dur)
L = len(hyp_tokens)
if len(toks_probs) >= L:
toks_probs = toks_probs[:L]
toks_durs = toks_durs[:L]
else:
pad = L - len(toks_probs)
toks_probs += [None]*pad
toks_durs += [None]*pad
arr = np.array([p for p in toks_probs if p is not None])
if arr.size:
low_t = float(np.quantile(arr, 0.15))
high_t = float(np.quantile(arr, 0.70))
else:
low_t, high_t = 0.5, 0.85
asr_token_conf = {i: {"prob": toks_probs[i], "duration_ms": toks_durs[i]} for i in range(L)}
return asr_token_conf, low_t, high_t
# =========================
# Confidence gate
# =========================
def gate_by_word_conf(base_decision: str, prob: float, sbert_sim: float,
is_short: bool, lev1: bool, duration_ms: float = None,
low_t: float = 0.6, high_t: float = 0.9, sbert_lo=0.60):
band = "mid"
if prob is not None:
if prob <= low_t: band = "low"
elif prob >= high_t: band = "high"
very_short = (duration_ms is not None and duration_ms < 120)
if band == "low":
if is_short and lev1: return 'ASR error (low p + short+lev1)'
if very_short: return 'ASR error (low p + very short)'
if sbert_sim >= sbert_lo: return 'ASR error (low p + semantic)'
return 'ASR error (low p)'
if band == "high":
return base_decision
return base_decision
# =========================
# Pair + main classifiers (tightened)
# =========================
def classify_pair(ref_w, hyp_w, bert_scores, phon_sim, lev1, short_word,
bert_thresh=0.75, max_bert=0.85):
# numbers equal
ref_num = to_numeric_value(ref_w)
hyp_num = to_numeric_value(hyp_w)
if (ref_num is not None) or (hyp_num is not None):
if (ref_num is not None) and (hyp_num is not None) and (ref_num == hyp_num):
return 'ASR error (numbers equal)'
# short+lev1
if short_word and lev1:
return 'ASR error (short+lev1)'
# semantic/phonetic
sbert_ok = bert_scores["sbert"] >= 0.80
avg_ok = bert_scores["avg"] >= bert_thresh
max_ok = (bert_scores["max"] > max_bert) and sbert_ok
disagree = (bert_scores.get("note") == "models_disagree")
if not disagree:
if ((phon_sim or lev1) and avg_ok) or max_ok:
return 'ASR error (semantic/phonetic)'
else:
if phon_sim or lev1:
if sbert_ok and avg_ok:
return 'ASR error (semantic/phonetic)'
else:
if bert_scores["sbert"] >= 0.80:
return 'ASR error (semantic)'
return 'Memorization error'
def classify_alignment_optimized(
aligned, ref_tokens, hyp_tokens,
bert_thresh=0.75, max_bert=0.85,
asr_token_conf=None, low_high=None,
replace_budget_tokens=None, # ุณูู ุงูุงุณุชุจุฏุงู
guard_note=None # ูุณู
ู
ุซู "off-topic"/"ok"/"budget_off"
):
# thresholds ู
ู ุงุญุชู
ุงูุงุช ุงูููู
ุงุช
if low_high is None:
if asr_token_conf:
probs = [v["prob"] for v in asr_token_conf.values() if v["prob"] is not None]
if probs:
low_t = float(np.quantile(probs, 0.15))
high_t = float(np.quantile(probs, 0.70))
else:
low_t, high_t = 0.5, 0.85
else:
low_t, high_t = 0.5, 0.85
else:
low_t, high_t = low_high
results, corrected_words = [], []
replaced_count = 0
for entry in aligned:
tag = entry['type']
i1, i2 = entry.get('ref_idx', (None, None))
j1, j2 = entry.get('hyp_idx', (None, None))
if tag == 'equal':
for ref_w, hyp_w in zip(entry['ref'], entry['hyp']):
results.append({'ASR_word': hyp_w, 'GT_word': ref_w, 'status': 'Correct', 'reason': '', 'used': hyp_w})
corrected_words.append(hyp_w)
elif tag in ['replace', 'delete', 'insert']:
max_len = max(len(entry['ref']), len(entry['hyp']))
for k in range(max_len):
ref_w = entry['ref'][k] if k < len(entry['ref']) else ''
hyp_w = entry['hyp'][k] if k < len(entry['hyp']) else ''
if not ref_w and not hyp_w:
continue
# similarities
phon_sim = phonetic_similarity(ref_w, hyp_w) if ref_w and hyp_w else False
lev1 = is_levenshtein_1(ref_w, hyp_w) if ref_w and hyp_w else False
bert_scores = multi_bert_similarity(ref_w, hyp_w) if ref_w and hyp_w else {"sbert":0,"marbert":0,"max":0,"avg":0}
short_word = bool(ref_w and hyp_w and max(len(ref_w), len(hyp_w)) <= 6)
# base status
if ref_w and hyp_w:
base_status = classify_pair(ref_w, hyp_w, bert_scores, phon_sim, lev1, short_word,
bert_thresh, max_bert)
elif hyp_w == '':
base_status = 'Missing (possible omission)'
elif ref_w == '':
base_status = 'Extra (possible ASR insertion)'
else:
base_status = 'Undefined Case'
# word-level confidence gate
word_prob = None; word_dur = None
if (j1 is not None) and (j2 is not None):
hyp_abs_idx = j1 + k
if asr_token_conf and hyp_abs_idx in asr_token_conf:
word_prob = asr_token_conf[hyp_abs_idx].get("prob")
word_dur = asr_token_conf[hyp_abs_idx].get("duration_ms")
final_status = base_status
if ref_w and hyp_w:
final_status = gate_by_word_conf(
base_decision=base_status, prob=word_prob,
sbert_sim=bert_scores["sbert"],
is_short=short_word, lev1=lev1,
duration_ms=word_dur,
low_t=low_t, high_t=high_t, sbert_lo=0.60
)
# choose token with budget
used = hyp_w
budget_info = ""
if ref_w and hyp_w:
if final_status.startswith("ASR error"):
if (replace_budget_tokens is None) or (replaced_count < replace_budget_tokens):
used = ref_w
replaced_count += 1
if replace_budget_tokens is not None:
budget_info = f", budget={replaced_count}/{replace_budget_tokens}"
else:
used = hyp_w
final_status += " [guard: budget reached]"
budget_info = f", budget={replaced_count}/{replace_budget_tokens}"
else:
used = hyp_w
elif hyp_w == '':
used = ''
elif ref_w == '':
used = hyp_w
reason = (f'Phonetic={phon_sim}, Lev1={lev1}, '
f'SBERT={bert_scores["sbert"]:.2f}, '
f'MARBERT={bert_scores["marbert"]:.2f}, '
f'MAX={bert_scores["max"]:.2f}, '
f'AVG={bert_scores["avg"]:.2f}, short={short_word}, '
f'prob={None if word_prob is None else round(word_prob,2)}, '
f'dur_ms={None if word_dur is None else int(word_dur)}, '
f'low_t={round(low_t,2)}, high_t={round(high_t,2)}')
if bert_scores.get("note"):
reason += f", note={bert_scores['note']}"
if guard_note:
reason += f", guard='{guard_note}'"
if budget_info:
reason += budget_info
results.append({
'ASR_word': hyp_w, 'GT_word': ref_w,
'status': final_status, 'reason': reason, 'used': used
})
if used:
corrected_words.append(used)
corrected_text = " ".join([w for w in corrected_words if w])
# ุฅุญุตุงุกุงุช ู
ุญููุฉ ู
ููุฏุฉ ููุชูุฑูุฑ
stats = {
"replacements_made": sum(1 for r in results
if r.get("used") and r.get("GT_word") and r["used"] == r["GT_word"]
and r.get("ASR_word") and r["ASR_word"] != r["GT_word"]),
"budget_reached_count": sum(1 for r in results if isinstance(r.get("status"), str) and "budget reached" in r["status"]),
"asr_error_count": sum(1 for r in results if isinstance(r.get("status"), str) and r["status"].startswith("ASR error")),
"memorization_error_count": sum(1 for r in results if r.get("status") == "Memorization error"),
"missing_count": sum(1 for r in results if r.get("status","").startswith("Missing")),
"extra_count": sum(1 for r in results if r.get("status","").startswith("Extra")),
"total_tokens": len(results)
}
return results, corrected_text, stats
# =========================
# ROUGE-L / WER-like / Guard
# =========================
def lcs_len(a, b):
m, n = len(a), len(b)
dp = [[0]*(n+1) for _ in range(m+1)]
for i in range(1, m+1):
ai = a[i-1]
for j in range(1, n+1):
if ai == b[j-1]:
dp[i][j] = dp[i-1][j-1] + 1
else:
dp[i][j] = dp[i-1][j] if dp[i-1][j] >= dp[i][j-1] else dp[i][j-1]
return dp[m][n]
def rouge_l_f1_tokens(ref_tokens, hyp_tokens, beta=1.2):
if not ref_tokens or not hyp_tokens:
return 0.0, 0.0, 0.0
lcs = lcs_len(ref_tokens, hyp_tokens)
prec = lcs / len(hyp_tokens)
rec = lcs / len(ref_tokens)
if prec == 0 and rec == 0:
return 0.0, 0.0, 0.0
f1 = ((1+beta**2) * prec * rec) / (rec + beta**2 * prec + 1e-12)
return float(f1), float(prec), float(rec)
def compute_wer_like(aligned, ref_tokens_len):
S = D = I = 0
for op in aligned:
if op['type'] == 'replace':
S += max(len(op['ref']), len(op['hyp']))
elif op['type'] == 'delete':
D += len(op['ref'])
elif op['type'] == 'insert':
I += len(op['hyp'])
N = max(ref_tokens_len, 1)
return (S + D + I) / N
def global_offtopic_guard(original_text, asr_text, ref_tokens, hyp_tokens, aligned, sbert_model):
sbert_sim_text = float(util.pytorch_cos_sim(
sbert_model.encode(_normalize_for_models(original_text), convert_to_tensor=True),
sbert_model.encode(_normalize_for_models(asr_text), convert_to_tensor=True)
))
rouge_f1, rouge_p, rouge_r = rouge_l_f1_tokens(ref_tokens, hyp_tokens)
equal_tokens = sum(len(op['ref']) for op in aligned if op['type'] == 'equal')
equal_ratio = equal_tokens / max(len(ref_tokens), 1)
wer = compute_wer_like(aligned, len(ref_tokens))
off_topic = ((sbert_sim_text < 0.70 and rouge_f1 < 0.45 and equal_ratio < 0.25) or (wer > 0.65))
L = len(hyp_tokens)
if off_topic:
budget = 0
elif sbert_sim_text < 0.80 or rouge_f1 < 0.55:
budget = int(0.15 * L)
else:
budget = int(0.40 * L)
metrics = {
"sbert_sim_text": round(sbert_sim_text, 3),
"rougeL_f1": round(rouge_f1, 3),
"rougeL_prec": round(rouge_p, 3),
"rougeL_rec": round(rouge_r, 3),
"equal_ratio": round(equal_ratio, 3),
"wer_like": round(wer, 3),
}
print(f"[GUARD] off_topic={off_topic}, budget={budget}, metrics={metrics}", flush=True)
return {"off_topic": off_topic, "budget_tokens": budget, "metrics": metrics}
# =========================
# Scores
# =========================
def literal_similarity(original, recited):
def norm(t):
t = re.sub(r'[ููููููููู]', '', t)
t = re.sub(r'[โโ",:ุุ.!()\[\]{}ุ\-โโ_]', ' ', t)
t = re.sub(r'\s+', ' ', t).strip()
return t
o = norm(original); r = norm(recited)
lev = textdistance.levenshtein.normalized_similarity(o, r)
ot = simple_tokenize(o); rt = simple_tokenize(r)
common = sum(1 for w1, w2 in zip(ot, rt) if w1 == w2)
word_overlap = common / max(len(ot), 1)
try:
import nltk.translate.bleu_score as bleu
bleu1 = bleu.sentence_bleu([ot], rt, weights=(1,0,0,0)) if (ot and rt) else 0.0
except Exception:
bleu1 = 0.0
final_score = 0.5*lev + 0.3*word_overlap + 0.2*bleu1
return {"levenshtein": round(lev,3), "word_overlap": round(word_overlap,3),
"bleu1": round(bleu1,3), "literal_score": round(final_score,3)}
def semantic_similarity(original, recited, use_marbert=FORCE_USE_MARBERT):
sbert_sim = float(util.pytorch_cos_sim(
_SBERT.encode(_normalize_for_models(original), convert_to_tensor=True),
_SBERT.encode(_normalize_for_models(recited), convert_to_tensor=True)
))
marbert_sim = marbert_cls_similarity(original, recited) if use_marbert else 0.0
return {"sbert_sim": round(sbert_sim,3), "marbert_sim": round(marbert_sim,3),
"semantic_score": round(max(sbert_sim, marbert_sim),3)}
# =========================
# Audio helper
# =========================
def ensure_audio_path(audio):
if isinstance(audio, str):
if not os.path.exists(audio):
raise FileNotFoundError(f"Audio path not found: {audio}")
return audio
if isinstance(audio, tuple) and len(audio) == 2:
data, sr = audio
if isinstance(data, np.ndarray):
tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
sf.write(tmp.name, data, sr)
return tmp.name
raise ValueError("Unsupported audio input format")
# =========================
# Pipeline (robust errors + logs)
# =========================
def transcribe_and_evaluate(audio, original_text, whisper_size=None,
compute_type=None, vad=True, use_marbert=True):
try:
if not original_text or not original_text.strip():
raise ValueError("Original text is empty.")
# Forced settings
whisper_size = FORCE_WHISPER_NAME
compute_type = FORCE_COMPUTE_TYPE
use_marbert = FORCE_USE_MARBERT
print(f"[RUN] whisper={whisper_size}, compute={compute_type}, marbert={use_marbert}", flush=True)
load_models(whisper_name=whisper_size, whisper_compute=compute_type, use_marbert=use_marbert)
audio_path = ensure_audio_path(audio)
print(f"[AUDIO] path={audio_path}", flush=True)
segments, info = _WHISPER.transcribe(audio_path, **ASR_OPTS)
segments = list(segments)
print(f"[ASR] segments={len(segments)}", flush=True)
# Build ASR text from words
words = []
for seg in segments:
for w in (seg.words or []):
tok = clean_ar_token(w.word)
if tok:
words.append(tok)
asr_text = " ".join(words)
# Tokens & alignment
ref_tokens = simple_tokenize(original_text)
hyp_tokens = simple_tokenize(asr_text)
aligned = align_texts(ref_tokens, hyp_tokens)
# Guard & budget
guard = global_offtopic_guard(original_text, asr_text, ref_tokens, hyp_tokens, aligned, _SBERT)
off_topic = guard["off_topic"]
guard_metrics = guard["metrics"]
if FORCE_BUDGET_MODE == "off":
budget_tokens = None
guard_note = "budget_off"
elif FORCE_BUDGET_MODE == "fixed":
budget_tokens = int(FIXED_BUDGET_TOKENS)
guard_note = f"budget_fixed_{budget_tokens}"
elif FORCE_BUDGET_MODE == "ratio":
budget_tokens = int(BUDGET_RATIO * len(hyp_tokens))
guard_note = f"budget_ratio_{BUDGET_RATIO}"
else:
budget_tokens = guard["budget_tokens"]
guard_note = "off-topic" if off_topic else "ok"
print(f"[BUDGET] mode={FORCE_BUDGET_MODE}, budget={budget_tokens}, note={guard_note}", flush=True)
# Word-level confidences
df_words = extract_word_conf_table(segments)
asr_token_conf, low_t, high_t = build_asr_token_conf(df_words, hyp_tokens)
print(f"[CONF] low_t={low_t:.3f}, high_t={high_t:.3f}", flush=True)
# Classification
results, corrected_text, local_stats = classify_alignment_optimized(
aligned, ref_tokens, hyp_tokens,
bert_thresh=0.75, max_bert=0.85,
asr_token_conf=asr_token_conf, low_high=(low_t, high_t),
replace_budget_tokens=budget_tokens,
guard_note=guard_note
)
# Scores
lit = literal_similarity(original_text, corrected_text)
sem = semantic_similarity(original_text, corrected_text, use_marbert=use_marbert)
# Extra global metrics for report
all_probs = df_words["prob"].dropna().tolist()
conf_summary = {
"num_words_with_prob": int(len(all_probs)),
"avg_prob": None if not all_probs else float(np.mean(all_probs)),
"p15": None if not all_probs else float(np.quantile(all_probs, 0.15)),
"p70": None if not all_probs else float(np.quantile(all_probs, 0.70)),
}
df = pd.DataFrame(results)
report = {
"requested": {"whisper_model": whisper_size, "compute_type": compute_type, "use_marbert": use_marbert},
"effective": {"whisper_model": whisper_size, "compute_type": compute_type, "use_marbert": use_marbert},
"guard": {"mode": FORCE_BUDGET_MODE, "off_topic": off_topic, "budget_tokens": None if budget_tokens is None else int(budget_tokens), **guard_metrics},
"local_stats": local_stats,
"confidence_summary": conf_summary,
"original_text": original_text,
"asr_text": asr_text,
"corrected_text": corrected_text,
"literal": lit,
"semantic": sem,
"low_t": float(low_t), "high_t": float(high_t),
}
return corrected_text, asr_text, json.dumps(report, ensure_ascii=False, indent=2), df
except Exception as e:
tb = traceback.format_exc()
print("ERROR in transcribe_and_evaluate:\n", tb, flush=True)
empty_df = pd.DataFrame([{"ASR_word":"","GT_word":"","status":"ERROR","reason":str(e),"used":""}])
err_json = json.dumps({"error": str(e), "traceback": tb}, ensure_ascii=False, indent=2)
gr.Warning(str(e))
return "", "", err_json, empty_df
def api_predict(audio, original_text, whisper_size=None, compute_type=None, vad=True, use_marbert=True):
corrected_text, asr_text, report_json, df = transcribe_and_evaluate(
audio, original_text, whisper_size, compute_type, vad, use_marbert
)
try:
return json.loads(report_json)
except Exception:
return {"error": "Failed to parse report_json."}
# =========================
# Gradio UI
# =========================
def build_ui():
with gr.Blocks(title="Samaali ASR Post-Processing", theme=gr.themes.Soft()) as demo:
gr.Markdown("## Samaali โ ASR Post-Processing (Whisper + Alignment + Confidence + Semantics)")
with gr.Row():
audio = gr.Audio(sources=["microphone","upload"], type="filepath", label="Audio")
original = gr.Textbox(lines=8, label="Original Text (Ground Truth)")
with gr.Row():
whisper_size = gr.Dropdown(choices=["large-v3"], value="large-v3", label="Whisper model size (forced)")
compute_type = gr.Dropdown(choices=["int8"], value="int8", label="compute_type (forced)")
vad = gr.Checkbox(value=True, label="VAD filter")
use_marbert = gr.Checkbox(value=True, label="Use MARBERT (forced)")
btn = gr.Button("Transcribe & Evaluate", variant="primary")
corrected = gr.Textbox(lines=6, label="Corrected Transcript (ASR errors restored)")
asr_out = gr.Textbox(lines=6, label="Raw ASR Transcript")
report = gr.JSON(label="Report (scores & thresholds)")
table = gr.Dataframe(headers=["ASR_word","GT_word","status","reason","used"],
label="Token-level Decisions", wrap=True)
btn.click(
fn=transcribe_and_evaluate,
inputs=[audio, original, whisper_size, compute_type, vad, use_marbert],
outputs=[corrected, asr_out, report, table],
api_name="evaluate"
)
gr.Button(visible=False).click(
fn=api_predict,
inputs=[audio, original, whisper_size, compute_type, vad, use_marbert],
outputs=gr.JSON(),
api_name="predict"
)
return demo
if __name__ == "__main__":
demo = build_ui()
demo.launch()
|