File size: 5,322 Bytes
48b78bf | 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 | """Canonical Bangla text normalizer = the paper's Stage-2 pipeline (single source of truth for
TTS input, CER reference, and training text). Order follows the paper:
1. digit conversion (Latin 0-9 -> Bengali) [paper step 1]
2. colon-numeral rewriting ("৪:৪৫" -> "৪ এর ৪৫") [paper step 2, from phonemizer.py]
3. numeral-to-word expansion via bnnumerizer [paper step 3] -- applied PER NUMBER
(bnnumerizer mangles the 2nd+ number in a string; per-number application is a bug-fix)
4. per-word Unicode normalization (bnunicodenormalizer) [paper step 4]
5. honorific expansion (paper attribution_dict, 12) [paper step 5]
6. whitespace collapse [paper step 6 = sentence seg + collapse]
Extensions beyond the paper (flagged, small): ordinals (৩য়->তৃতীয়), title abbreviations (ড./ডা./মো.),
thousands-comma stripping. The paper lists domain abbreviations as a known gap.
Use `for_tts(text)` before synthesis and `for_cer(text)` for reference/hypothesis before CER."""
import re, unicodedata
from bnnumerizer import numerize
try:
from bnunicodenormalizer import Normalizer
_bnorm = Normalizer()
except Exception:
_bnorm = None
# --- paper attribution_dict (phonemizer.py) ---
HONORIFICS = {
"সাঃ": "সাল্লাল্লাহু আলাইহি ওয়া সাল্লাম", "রাঃ": "রাদিআল্লাহু আনহু", "রহঃ": "রাহিমাহুল্লাহ",
"হাঃ": "হাফিজাহুল্লাহ", "মাঃ": "মাদ্দাজিল্লুহু", "দাঃ বাঃ": "দামাত বারাকাতুহুম",
"আঃ": "আলাইহিস সালাম", "জিঃ": "জিল্লুল্লাহ", "কুঃ সিঃ": "কুদ্দিসা সিরুহু",
"আলঃ": "আলাইহিমুস সালাম", "রাযিঃ": "রাযিআল্লাহু আনহু", "মুহঃ": "মুহতারাম",
}
# --- extensions (not in paper; keep explicit) ---
ORDINALS = {"১ম":"প্রথম","২য়":"দ্বিতীয়","৩য়":"তৃতীয়","৪র্থ":"চতুর্থ","৫ম":"পঞ্চম","৬ষ্ঠ":"ষষ্ঠ","৭ম":"সপ্তম","৮ম":"অষ্টম","৯ম":"নবম","১০ম":"দশম"}
TITLES = {"ড.":"ডক্টর","ডা.":"ডাক্তার","মো.":"মোহাম্মদ","মোঃ":"মোহাম্মদ","প্রফ.":"প্রফেসর","ইঞ্জি.":"ইঞ্জিনিয়ার","অ্যাড.":"অ্যাডভোকেট"}
EN2BN = {"laptop":"ল্যাপটপ","monday":"মানডে","email":"ইমেইল","password":"পাসওয়ার্ড","software":"সফটওয়্যার",
"performance":"পারফরম্যান্স","team":"টিম","project":"প্রজেক্ট","online":"অনলাইন"}
_LAT2BN = str.maketrans("0123456789", "০১২৩৪৫৬৭৮৯")
_NUM = re.compile(r"[০-৯]+(?:[.][০-৯]+)*")
_COLON = re.compile(r"([০-৯]+):([০-৯]+)")
def _unicode_words(s):
if _bnorm is None: return s
out = []
for w in s.split():
try: r = _bnorm(w); out.append(r["normalized"] or w)
except Exception: out.append(w)
return " ".join(out)
def for_tts(text: str) -> str:
"""Paper-order normalization producing the string the model should speak."""
t = str(text)
for k, v in sorted(TITLES.items(), key=lambda kv: -len(kv[0])): t = t.replace(k, v) # ext
t = t.translate(_LAT2BN) # 1
t = _COLON.sub(lambda m: f"{m.group(1)} এর {m.group(2)}", t) # 2
for k, v in sorted(ORDINALS.items(), key=lambda kv: -len(kv[0])): t = t.replace(k, v) # ext
t = t.replace(",", "") # ext (bnnumerizer breaks on ১,২৫০)
t = _NUM.sub(lambda m: numerize(m.group(0)), t) # 3 (per number)
t = _unicode_words(t) # 4
for k, v in sorted(HONORIFICS.items(), key=lambda kv: -len(kv[0])): t = t.replace(k, _unicode_words(v)) # 5
return re.sub(r"\s+", " ", t).strip() # 6
def for_cer(text: str) -> str:
"""Same expansions as for_tts, plus English->Bengali map, NFC, ZW strip, punctuation/space removal."""
t = re.sub(r"[A-Za-z]+", lambda m: EN2BN.get(m.group(0).lower(), m.group(0)), str(text))
t = for_tts(t)
t = unicodedata.normalize("NFC", t).replace("", "").replace("", "")
return re.sub(r"[।,\.!?;:\-—\s\"'‘’“”()]", "", t)
if __name__ == "__main__":
for s in ["নবী (সাঃ) বলেছেন","হযরত আলী (রাঃ) ও ইমাম আবু হানিফা (রহঃ)","সভা ৪:৪৫ এ শুরু","ড. ইউনূস ৩য় তলায় ১,২৫০ টাকা দিলেন","1966 সালে 400 টি বাঘ"]:
print(repr(s), "->", for_tts(s))
|