from __future__ import annotations from app.text import levenshtein # Romanized Hindi function words. These sound like the Devanagari, so they must # stay in Hindi rather than be copied from the English pass. _ROMAN_HINDI = frozenset( { "hai", "hain", "nahi", "nahin", "nhi", "kya", "kyun", "kyu", "mein", "main", "hum", "ham", "aap", "ap", "ye", "yeh", "yah", "woh", "wo", "vo", "aur", "toh", "ka", "ki", "ke", "se", "par", "pe", "bhi", "jo", "jab", "tab", "ab", "phir", "fir", "lekin", "magar", "kyunki", "bahut", "thoda", "accha", "achha", "theek", "thik", "haan", "han", "mat", "kar", "karna", "karenge", "kiya", "hua", "hui", "hoga", "hogi", "tha", "thi", "ho", "hota", "hoti", "raha", "rahe", "rahi", "wala", "wali", "wale", "iska", "uska", "unka", "mera", "teri", "tera", "hamara", "kyon", "agar", "ya", "ek", "do", "kuch", "sab", "bhai", "ji", "ha", } ) _SIGNS = { "अ": "a", "आ": "aa", "इ": "i", "ई": "ii", "उ": "u", "ऊ": "uu", "ऋ": "ri", "ए": "e", "ऐ": "ai", "ओ": "o", "औ": "au", "ा": "a", "ि": "i", "ी": "ii", "ु": "u", "ू": "uu", "ृ": "ri", "े": "e", "ै": "ai", "ो": "o", "ौ": "au", "ॉ": "o", "क": "k", "ख": "kh", "ग": "g", "घ": "gh", "ङ": "ng", "च": "ch", "छ": "chh", "ज": "j", "झ": "jh", "ञ": "ny", "ट": "t", "ठ": "th", "ड": "d", "ढ": "dh", "ण": "n", "त": "t", "थ": "th", "द": "d", "ध": "dh", "न": "n", "प": "p", "फ": "ph", "ब": "b", "भ": "bh", "म": "m", "य": "y", "र": "r", "ल": "l", "व": "v", "श": "sh", "ष": "sh", "स": "s", "ह": "h", "क़": "k", "ख़": "kh", "ग़": "g", "ज़": "z", "ड़": "d", "ढ़": "dh", "फ़": "f", "य़": "y", "ं": "n", "ँ": "n", "ः": "h", "़": "", "्": "", "ॐ": "", } def blend_words(hindi_words: list[dict], english_words: list[dict], pause: float = 0.7) -> list[dict]: chosen = _choose(hindi_words, english_words) return _group(chosen, pause) def _choose(hindi_words: list[dict], english_words: list[dict]) -> list[dict]: chosen: list[dict] = [] for word in hindi_words: cleaned = _keep_spoken_scripts(str(word.get("text") or "")) if not _has_letter(cleaned): continue match = _best_overlap(word, english_words) text = cleaned marker = None if match is not None: english = _latin_word(str(match.get("text") or "")) if english and _prefer_english(cleaned, english): text = english marker = id(match) if marker is not None and chosen and chosen[-1].get("_en") == marker: chosen[-1]["end"] = float(word.get("end") or chosen[-1]["end"]) continue item = { "start": float(word.get("start") or 0), "end": float(word.get("end") or 0), "text": text, } if marker is not None: item["_en"] = marker chosen.append(item) for item in chosen: item.pop("_en", None) return chosen def _prefer_english(hindi: str, english: str) -> bool: key = english.lower() if key in _ROMAN_HINDI: return False if not _sounds_like(hindi, english): return False return True def _sounds_like(hindi: str, english: str) -> bool: spoken = "".join(char for char in english.lower() if char.isalpha()) heard = _latinize(hindi) if not spoken or not heard: return False if _close(_skeleton(spoken), _skeleton(heard)): return True return len(spoken) <= 4 and _close(spoken, heard, limit=0.45) def _close(left: str, right: str, limit: float = 0.34) -> bool: if not left or not right: return False if len(left) < 3 or len(right) < 3: return left == right and len(left) >= 2 if left == right: return True return levenshtein(left, right) / max(len(left), len(right)) <= limit def _latinize(text: str) -> str: parts: list[str] = [] for char in text: if char in _SIGNS: parts.append(_SIGNS[char]) elif "a" <= char.lower() <= "z": parts.append(char.lower()) return "".join(parts) def _skeleton(text: str) -> str: return "".join(char for char in text if char not in "aeiou") def _best_overlap(word: dict, others: list[dict]) -> dict | None: start = float(word.get("start") or 0) end = float(word.get("end") or start) if end < start: end = start best = None best_score = 0.0 for other in others: other_start = float(other.get("start") or 0) other_end = float(other.get("end") or other_start) overlap = min(end, other_end) - max(start, other_start) if overlap <= 0: continue shorter = max(0.05, min(end - start, other_end - other_start)) score = overlap / shorter if score > best_score: best_score = score best = other if best_score < 0.4: return None return best def _keep_spoken_scripts(text: str) -> str: kept: list[str] = [] for char in text: if char.isascii() or "\u0900" <= char <= "\u097f" or char.isspace(): if char == "\u0950": continue kept.append(char) elif char in "।॥": kept.append(char) return " ".join("".join(kept).split()) def _latin_word(text: str) -> str: letters = [char for char in text if char.isascii() and (char.isalnum() or char in "-'")] word = "".join(letters).strip("-'") if not any(char.isalpha() for char in word): return "" return word def _has_letter(text: str) -> bool: return any(char.isalpha() or char.isdigit() for char in text) def _has_devanagari(text: str) -> bool: return any("\u0900" <= char <= "\u097f" for char in text) def _group(words: list[dict], pause: float) -> list[dict]: if not words: return [] groups = [[words[0]]] for word in words[1:]: gap = float(word["start"]) - float(groups[-1][-1]["end"]) if gap >= pause: groups.append([word]) else: groups[-1].append(word) chunks = [] for group in groups: text = " ".join(item["text"] for item in group).strip() if not text: continue chunks.append({"start": group[0]["start"], "end": group[-1]["end"], "text": text}) return chunks