Upload 2 files

text/japanese.py

@@ -1,704 +1,704 @@
-# Convert Japanese text to phonemes which is
-# compatible with Julius https://github.com/julius-speech/segmentation-kit
-import re
-import unicodedata
-
-from transformers import AutoTokenizer
-
-from text import punctuation, symbols
-
-try:
-    import MeCab
-except ImportError as e:
-    raise ImportError("Japanese requires mecab-python3 and unidic-lite.") from e
-from num2words import num2words
-
-_CONVRULES = [
-    # Conversion of 2 letters
-    "アァ/ a a",
-    "イィ/ i i",
-    "イェ/ i e",
-    "イャ/ y a",
-    "ウゥ/ u:",
-    "エェ/ e e",
-    "オォ/ o:",
-    "カァ/ k a:",
-    "キィ/ k i:",
-    "クゥ/ k u:",
-    "クャ/ ky a",
-    "クュ/ ky u",
-    "クョ/ ky o",
-    "ケェ/ k e:",
-    "コォ/ k o:",
-    "ガァ/ g a:",
-    "ギィ/ g i:",
-    "グゥ/ g u:",
-    "グャ/ gy a",
-    "グュ/ gy u",
-    "グョ/ gy o",
-    "ゲェ/ g e:",
-    "ゴォ/ g o:",
-    "サァ/ s a:",
-    "シィ/ sh i:",
-    "スゥ/ s u:",
-    "スャ/ sh a",
-    "スュ/ sh u",
-    "スョ/ sh o",
-    "セェ/ s e:",
-    "ソォ/ s o:",
-    "ザァ/ z a:",
-    "ジィ/ j i:",
-    "ズゥ/ z u:",
-    "ズャ/ zy a",
-    "ズュ/ zy u",
-    "ズョ/ zy o",
-    "ゼェ/ z e:",
-    "ゾォ/ z o:",
-    "タァ/ t a:",
-    "チィ/ ch i:",
-    "ツァ/ ts a",
-    "ツィ/ ts i",
-    "ツゥ/ ts u:",
-    "ツャ/ ch a",
-    "ツュ/ ch u",
-    "ツョ/ ch o",
-    "ツェ/ ts e",
-    "ツォ/ ts o",
-    "テェ/ t e:",
-    "トォ/ t o:",
-    "ダァ/ d a:",
-    "ヂィ/ j i:",
-    "ヅゥ/ d u:",
-    "ヅャ/ zy a",
-    "ヅュ/ zy u",
-    "ヅョ/ zy o",
-    "デェ/ d e:",
-    "ドォ/ d o:",
-    "ナァ/ n a:",
-    "ニィ/ n i:",
-    "ヌゥ/ n u:",
-    "ヌャ/ ny a",
-    "ヌュ/ ny u",
-    "ヌョ/ ny o",
-    "ネェ/ n e:",
-    "ノォ/ n o:",
-    "ハァ/ h a:",
-    "ヒィ/ h i:",
-    "フゥ/ f u:",
-    "フャ/ hy a",
-    "フュ/ hy u",
-    "フョ/ hy o",
-    "ヘェ/ h e:",
-    "ホォ/ h o:",
-    "バァ/ b a:",
-    "ビィ/ b i:",
-    "ブゥ/ b u:",
-    "フャ/ hy a",
-    "ブュ/ by u",
-    "フョ/ hy o",
-    "ベェ/ b e:",
-    "ボォ/ b o:",
-    "パァ/ p a:",
-    "ピィ/ p i:",
-    "プゥ/ p u:",
-    "プャ/ py a",
-    "プュ/ py u",
-    "プョ/ py o",
-    "ペェ/ p e:",
-    "ポォ/ p o:",
-    "マァ/ m a:",
-    "ミィ/ m i:",
-    "ムゥ/ m u:",
-    "ムャ/ my a",
-    "ムュ/ my u",
-    "ムョ/ my o",
-    "メェ/ m e:",
-    "モォ/ m o:",
-    "ヤァ/ y a:",
-    "ユゥ/ y u:",
-    "ユャ/ y a:",
-    "ユュ/ y u:",
-    "ユョ/ y o:",
-    "ヨォ/ y o:",
-    "ラァ/ r a:",
-    "リィ/ r i:",
-    "ルゥ/ r u:",
-    "ルャ/ ry a",
-    "ルュ/ ry u",
-    "ルョ/ ry o",
-    "レェ/ r e:",
-    "ロォ/ r o:",
-    "ワァ/ w a:",
-    "ヲォ/ o:",
-    "ディ/ d i",
-    "デェ/ d e:",
-    "デャ/ dy a",
-    "デュ/ dy u",
-    "デョ/ dy o",
-    "ティ/ t i",
-    "テェ/ t e:",
-    "テャ/ ty a",
-    "テュ/ ty u",
-    "テョ/ ty o",
-    "スィ/ s i",
-    "ズァ/ z u a",
-    "ズィ/ z i",
-    "ズゥ/ z u",
-    "ズャ/ zy a",
-    "ズュ/ zy u",
-    "ズョ/ zy o",
-    "ズェ/ z e",
-    "ズォ/ z o",
-    "キャ/ ky a",
-    "キュ/ ky u",
-    "キョ/ ky o",
-    "シャ/ sh a",
-    "シュ/ sh u",
-    "シェ/ sh e",
-    "ショ/ sh o",
-    "チャ/ ch a",
-    "チュ/ ch u",
-    "チェ/ ch e",
-    "チョ/ ch o",
-    "トゥ/ t u",
-    "トャ/ ty a",
-    "トュ/ ty u",
-    "トョ/ ty o",
-    "ドァ/ d o a",
-    "ドゥ/ d u",
-    "ドャ/ dy a",
-    "ドュ/ dy u",
-    "ドョ/ dy o",
-    "ドォ/ d o:",
-    "ニャ/ ny a",
-    "ニュ/ ny u",
-    "ニョ/ ny o",
-    "ヒャ/ hy a",
-    "ヒュ/ hy u",
-    "ヒョ/ hy o",
-    "ミャ/ my a",
-    "ミュ/ my u",
-    "ミョ/ my o",
-    "リャ/ ry a",
-    "リュ/ ry u",
-    "リョ/ ry o",
-    "ギャ/ gy a",
-    "ギュ/ gy u",
-    "ギョ/ gy o",
-    "ヂェ/ j e",
-    "ヂャ/ j a",
-    "ヂュ/ j u",
-    "ヂョ/ j o",
-    "ジェ/ j e",
-    "ジャ/ j a",
-    "ジュ/ j u",
-    "ジョ/ j o",
-    "ビャ/ by a",
-    "ビュ/ by u",
-    "ビョ/ by o",
-    "ピャ/ py a",
-    "ピュ/ py u",
-    "ピョ/ py o",
-    "ウァ/ u a",
-    "ウィ/ w i",
-    "ウェ/ w e",
-    "ウォ/ w o",
-    "ファ/ f a",
-    "フィ/ f i",
-    "フゥ/ f u",
-    "フャ/ hy a",
-    "フュ/ hy u",
-    "フョ/ hy o",
-    "フェ/ f e",
-    "フォ/ f o",
-    "ヴァ/ b a",
-    "ヴィ/ b i",
-    "ヴェ/ b e",
-    "ヴォ/ b o",
-    "ヴュ/ by u",
-    "アー/ a:",
-    "イー/ i:",
-    "ウー/ u:",
-    "エー/ e:",
-    "オー/ o:",
-    "カー/ k a:",
-    "キー/ k i:",
-    "クー/ k u:",
-    "ケー/ k e:",
-    "コー/ k o:",
-    "サー/ s a:",
-    "シー/ sh i:",
-    "スー/ s u:",
-    "セー/ s e:",
-    "ソー/ s o:",
-    "ター/ t a:",
-    "チー/ ch i:",
-    "ツー/ ts u:",
-    "テー/ t e:",
-    "トー/ t o:",
-    "ナー/ n a:",
-    "ニー/ n i:",
-    "��ー/ n u:",
-    "ネー/ n e:",
-    "ノー/ n o:",
-    "ハー/ h a:",
-    "ヒー/ h i:",
-    "フー/ f u:",
-    "ヘー/ h e:",
-    "ホー/ h o:",
-    "マー/ m a:",
-    "ミー/ m i:",
-    "ムー/ m u:",
-    "メー/ m e:",
-    "モー/ m o:",
-    "ラー/ r a:",
-    "リー/ r i:",
-    "ルー/ r u:",
-    "レー/ r e:",
-    "ロー/ r o:",
-    "ガー/ g a:",
-    "ギー/ g i:",
-    "グー/ g u:",
-    "ゲー/ g e:",
-    "ゴー/ g o:",
-    "ザー/ z a:",
-    "ジー/ j i:",
-    "ズー/ z u:",
-    "ゼー/ z e:",
-    "ゾー/ z o:",
-    "ダー/ d a:",
-    "ヂー/ j i:",
-    "ヅー/ z u:",
-    "デー/ d e:",
-    "ドー/ d o:",
-    "バー/ b a:",
-    "ビー/ b i:",
-    "ブー/ b u:",
-    "ベー/ b e:",
-    "ボー/ b o:",
-    "パー/ p a:",
-    "ピー/ p i:",
-    "プー/ p u:",
-    "ペー/ p e:",
-    "ポー/ p o:",
-    "ヤー/ y a:",
-    "ユー/ y u:",
-    "ヨー/ y o:",
-    "ワー/ w a:",
-    "ヰー/ i:",
-    "ヱー/ e:",
-    "ヲー/ o:",
-    "ヴー/ b u:",
-    # Conversion of 1 letter
-    "ア/ a",
-    "イ/ i",
-    "ウ/ u",
-    "エ/ e",
-    "オ/ o",
-    "カ/ k a",
-    "キ/ k i",
-    "ク/ k u",
-    "ケ/ k e",
-    "コ/ k o",
-    "サ/ s a",
-    "シ/ sh i",
-    "ス/ s u",
-    "セ/ s e",
-    "ソ/ s o",
-    "タ/ t a",
-    "チ/ ch i",
-    "ツ/ ts u",
-    "テ/ t e",
-    "ト/ t o",
-    "ナ/ n a",
-    "ニ/ n i",
-    "ヌ/ n u",
-    "ネ/ n e",
-    "ノ/ n o",
-    "ハ/ h a",
-    "ヒ/ h i",
-    "フ/ f u",
-    "ヘ/ h e",
-    "ホ/ h o",
-    "マ/ m a",
-    "ミ/ m i",
-    "ム/ m u",
-    "メ/ m e",
-    "モ/ m o",
-    "ラ/ r a",
-    "リ/ r i",
-    "ル/ r u",
-    "レ/ r e",
-    "ロ/ r o",
-    "ガ/ g a",
-    "ギ/ g i",
-    "グ/ g u",
-    "ゲ/ g e",
-    "ゴ/ g o",
-    "ザ/ z a",
-    "ジ/ j i",
-    "ズ/ z u",
-    "ゼ/ z e",
-    "ゾ/ z o",
-    "ダ/ d a",
-    "ヂ/ j i",
-    "ヅ/ z u",
-    "デ/ d e",
-    "ド/ d o",
-    "バ/ b a",
-    "ビ/ b i",
-    "ブ/ b u",
-    "ベ/ b e",
-    "ボ/ b o",
-    "パ/ p a",
-    "ピ/ p i",
-    "プ/ p u",
-    "ペ/ p e",
-    "ポ/ p o",
-    "ヤ/ y a",
-    "ユ/ y u",
-    "ヨ/ y o",
-    "ワ/ w a",
-    "ヰ/ i",
-    "ヱ/ e",
-    "ヲ/ o",
-    "ン/ N",
-    "ッ/ q",
-    "ヴ/ b u",
-    "ー/:", #这个不起作用
-    # Try converting broken text
-    "ァ/ a",
-    "ィ/ i",
-    "ゥ/ u",
-    "ェ/ e",
-    "ォ/ o",
-    "ヮ/ w a",
-    "ォ/ o",
-    # Symbols
-    "、/ ,",
-    "。/ .",
-    "！/ !",
-    "？/ ?",
-    "・/ ,",
-]
-
-_COLON_RX = re.compile(":+")
-_REJECT_RX = re.compile("[^ a-zA-Z:,.?]")
-
-
-def _makerulemap():
-    l = [tuple(x.split("/")) for x in _CONVRULES]
-    return tuple({k: v for k, v in l if len(k) == i} for i in (1, 2))
-
-
-_RULEMAP1, _RULEMAP2 = _makerulemap()
-
-
-def kata2phoneme(text: str) -> str:
-    """Convert katakana text to phonemes."""
-    text = text.strip()
-    res = []
-    while text:
-        if len(text) >= 2:
-            x = _RULEMAP2.get(text[:2])
-            if x is not None:
-                text = text[2:]
-                res += x.split(" ")[1:]
-                continue
-        x = _RULEMAP1.get(text[0])
-        if x is not None:
-            text = text[1:]
-            res += x.split(" ")[1:]
-            continue
-        res.append(text[0])
-        text = text[1:]
-    # res = _COLON_RX.sub(":", res)
-    return res
-
-
-_KATAKANA = "".join(chr(ch) for ch in range(ord("ァ"), ord("ン") + 1))
-_HIRAGANA = "".join(chr(ch) for ch in range(ord("ぁ"), ord("ん") + 1))
-_HIRA2KATATRANS = str.maketrans(_HIRAGANA, _KATAKANA)
-
-
-def hira2kata(text: str) -> str:
-    text = text.translate(_HIRA2KATATRANS)
-    return text.replace("う゛", "ヴ")
-
-
-_SYMBOL_TOKENS = set(list("・、。？！"))
-_NO_YOMI_TOKENS = set(list("「」『』―（）［］[]"))
-_TAGGER = MeCab.Tagger()
-
-
-def text2kata(text: str) -> str:
-    parsed = _TAGGER.parse(text)
-    res = []
-    for line in parsed.split("\n"):
-        if line == "EOS":
-            break
-        parts = line.split("\t")
-
-        word, yomi = parts[0], parts[1]
-        if yomi:
-            res.append(yomi)
-        else:
-            if word in _SYMBOL_TOKENS:
-                res.append(word)
-            elif word in ("っ", "ッ"):
-                res.append("ッ")
-            elif word in _NO_YOMI_TOKENS:
-                pass
-            else:
-                res.append(word)
-    return hira2kata("".join(res))
-
-
-def text2sep_kata(text: str) -> (list, list):
-    parsed = _TAGGER.parse(text)
-    res = []
-    sep = []
-    for line in parsed.split("\n"):
-        if line == "EOS":
-            break
-        parts = line.split("\t")
-
-        word, yomi = parts[0], parts[1]
-        if yomi:
-            res.append(yomi)
-        else:
-            if word in _SYMBOL_TOKENS:
-                res.append(word)
-            elif word in ("っ", "ッ"):
-                res.append("ッ")
-            elif word in _NO_YOMI_TOKENS:
-                pass
-            else:
-                res.append(word)
-        sep.append(word)
-    return sep, [hira2kata(i) for i in res]
-
-
-_ALPHASYMBOL_YOMI = {
-    "#": "シャープ",
-    "%": "パーセント",
-    "&": "アンド",
-    "+": "プラス",
-    "-": "マイナス",
-    ":": "コロン",
-    ";": "セミコロン",
-    "<": "小なり",
-    "=": "イコール",
-    ">": "大なり",
-    "@": "アット",
-    "a": "エー",
-    "b": "ビー",
-    "c": "シー",
-    "d": "ディー",
-    "e": "イー",
-    "f": "エフ",
-    "g": "ジー",
-    "h": "エイチ",
-    "i": "アイ",
-    "j": "ジェー",
-    "k": "ケー",
-    "l": "エル",
-    "m": "エム",
-    "n": "エヌ",
-    "o": "オー",
-    "p": "ピー",
-    "q": "キュー",
-    "r": "アール",
-    "s": "エス",
-    "t": "ティー",
-    "u": "ユー",
-    "v": "ブイ",
-    "w": "ダブリュー",
-    "x": "エックス",
-    "y": "ワイ",
-    "z": "ゼット",
-    "α": "アルファ",
-    "β": "ベータ",
-    "γ": "ガンマ",
-    "δ": "デルタ",
-    "ε": "イプシロン",
-    "ζ": "ゼータ",
-    "η": "イータ",
-    "θ": "シータ",
-    "ι": "イオタ",
-    "κ": "カッパ",
-    "λ": "ラムダ",
-    "μ": "ミュー",
-    "ν": "ニュー",
-    "ξ": "クサイ",
-    "ο": "オミクロン",
-    "π": "パイ",
-    "ρ": "ロー",
-    "σ": "シグマ",
-    "τ": "タウ",
-    "υ": "ウプシロン",
-    "φ": "ファイ",
-    "χ": "カイ",
-    "ψ": "プサイ",
-    "ω": "オメガ",
-}
-
-
-_NUMBER_WITH_SEPARATOR_RX = re.compile("[0-9]{1,3}(,[0-9]{3})+")
-_CURRENCY_MAP = {"$": "ドル", "¥": "円", "£": "ポンド", "€": "ユーロ"}
-_CURRENCY_RX = re.compile(r"([$¥£€])([0-9.]*[0-9])")
-_NUMBER_RX = re.compile(r"[0-9]+(\.[0-9]+)?")
-
-
-def japanese_convert_numbers_to_words(text: str) -> str:
-    res = _NUMBER_WITH_SEPARATOR_RX.sub(lambda m: m[0].replace(",", ""), text)
-    res = _CURRENCY_RX.sub(lambda m: m[2] + _CURRENCY_MAP.get(m[1], m[1]), res)
-    res = _NUMBER_RX.sub(lambda m: num2words(m[0], lang="ja"), res)
-    return res
-
-
-def japanese_convert_alpha_symbols_to_words(text: str) -> str:
-    return "".join([_ALPHASYMBOL_YOMI.get(ch, ch) for ch in text.lower()])
-
-
-def japanese_text_to_phonemes(text: str) -> str:
-    """Convert Japanese text to phonemes."""
-    res = unicodedata.normalize("NFKC", text)
-    res = japanese_convert_numbers_to_words(res)
-    # res = japanese_convert_alpha_symbols_to_words(res)
-    res = text2kata(res)
-    res = kata2phoneme(res)
-    return res
-
-
-def is_japanese_character(char):
-    # 定义日语文字系统的 Unicode 范围
-    japanese_ranges = [
-        (0x3040, 0x309F),  # 平假名
-        (0x30A0, 0x30FF),  # 片假名
-        (0x4E00, 0x9FFF),  # 汉字 (CJK Unified Ideographs)
-        (0x3400, 0x4DBF),  # 汉字扩展 A
-        (0x20000, 0x2A6DF),  # 汉字扩展 B
-        # 可以根据需要添加其他汉字扩展范围
-    ]
-
-    # 将字符的 Unicode 编码转换为整数
-    char_code = ord(char)
-
-    # 检查字符是否在任何一个日语范围内
-    for start, end in japanese_ranges:
-        if start <= char_code <= end:
-            return True
-
-    return False
-
-
-rep_map = {
-    "：": ",",
-    "；": ",",
-    "，": ",",
-    "。": ".",
-    "！": "!",
-    "？": "?",
-    "\n": ".",
-    "·": ",",
-    "、": ",",
-    "…": "...",
-}
-
-
-def replace_punctuation(text):
-    pattern = re.compile("|".join(re.escape(p) for p in rep_map.keys()))
-
-    replaced_text = pattern.sub(lambda x: rep_map[x.group()], text)
-
-    replaced_text = re.sub(
-        r"[^\u3040-\u309F\u30A0-\u30FF\u4E00-\u9FFF\u3400-\u4DBF"
-        + "".join(punctuation)
-        + r"]+",
-        "",
-        replaced_text,
-    )
-
-    return replaced_text
-
-
-def text_normalize(text):
-    res = unicodedata.normalize("NFKC", text)
-    res = japanese_convert_numbers_to_words(res)
-    # res = "".join([i for i in res if is_japanese_character(i)])
-    res = replace_punctuation(res)
-    return res
-
-
-def distribute_phone(n_phone, n_word):
-    phones_per_word = [0] * n_word
-    for task in range(n_phone):
-        min_tasks = min(phones_per_word)
-        min_index = phones_per_word.index(min_tasks)
-        phones_per_word[min_index] += 1
-    return phones_per_word
-
-
-tokenizer = AutoTokenizer.from_pretrained("./bert/bert-base-japanese-v3")
-
-
-def g2p(norm_text):
-    sep_text, sep_kata = text2sep_kata(norm_text)
-    sep_tokenized = [tokenizer.tokenize(i) for i in sep_text]
-    sep_phonemes = [kata2phoneme(i) for i in sep_kata]
-    # 异常处理，MeCab不认识的词的话会一路传到这里来，然后炸掉。目前来看只有那些超级稀有的生僻词会出现这种情况
-    for i in sep_phonemes:
-        for j in i:
-            assert j in symbols, (sep_text, sep_kata, sep_phonemes)
-
-    word2ph = []
-    for token, phoneme in zip(sep_tokenized, sep_phonemes):
-        phone_len = len(phoneme)
-        word_len = len(token)
-
-        aaa = distribute_phone(phone_len, word_len)
-        word2ph += aaa
-    phones = ["_"] + [j for i in sep_phonemes for j in i] + ["_"]
-    tones = [0 for i in phones]
-    word2ph = [1] + word2ph + [1]
-    return phones, tones, word2ph
-
-if __name__ == "__main__":
-    tokenizer = AutoTokenizer.from_pretrained("./bert/bert-base-japanese-v3")
-    text = "だったら私、スズカさんと同じチームに入りたいです！ スズカさんの走りを毎日近くで、なんなら真横から見ていたいので！"
-    #print(_TAGGER.parse(text))
-   # nodes = [{"surface": "こんにちは", "pos": "感動詞:*:*:*", "pron": "コンニチワ", "c_type": "*", "c_form": "*", "accent_type": 0, "accent_con_type": "-1", "chain_flag": -1}]
-    nodes = [{"surface":"こんにちは","pron": "コンニチワ","pos": "感動詞:*:*:*",}]
-    from text.japanese_bert import get_bert_feature
-    import pyopenjtalk
-    from marine.predict import Predictor
-    from marine.utils.openjtalk_util import convert_njd_feature_to_marine_feature
-    text = text_normalize(text)
-    NJD_NODES = pyopenjtalk.run_frontend(text)
-    predictor = Predictor()
- #   important_info = [{"string":i["string"],"pron":i["pron"],"acc":i["acc"]}for i in pyopenjtalk.estimate_accent(NJD_NODES)]
-    print(text)
-    
-    marine_feature = convert_njd_feature_to_marine_feature(NJD_NODES)
-    results = predictor.predict([marine_feature])
-    for mora,acc in zip(results["mora"][0],results["accent_status"][0]):
-        print(f"{mora}:{acc}")
-    # for i in pyopenjtalk.estimate_accent(NJD_NODES):
-    #     print(f"{i['string']}:{i['pron']}:{i['acc']}")
-#     info = pyopenjtalk.extract_fullcontext(text,run_marine=True)
-#     info_nomarine = pyopenjtalk.extract_fullcontext(text,run_marine=False)
-#    # nodes = pyopenjtalk
-#    # print(info)
-#     for i,j in zip(info,info_nomarine):
-#         print(i)
-#         print(j)
-#         print("\n")
-    # predictor = Predictor()
-    #print(pyopenjtalk.estimate_accent(text))
-    # output = predictor.predict([nodes],accent_represent_mode="high_low")
-    #print(output)
-    # phones, tones, word2ph = g2p(text)
-    # bert = get_bert_feature(text, word2ph)
-
-    # print(phones, tones, word2ph, bert.shape)
+# Convert Japanese text to phonemes which is
+# compatible with Julius https://github.com/julius-speech/segmentation-kit
+import re
+import unicodedata
+
+from transformers import AutoTokenizer
+
+from text import punctuation, symbols
+
+try:
+    import MeCab
+except ImportError as e:
+    raise ImportError("Japanese requires mecab-python3 and unidic-lite.") from e
+from num2words import num2words
+
+_CONVRULES = [
+    # Conversion of 2 letters
+    "アァ/ a a",
+    "イィ/ i i",
+    "イェ/ i e",
+    "イャ/ y a",
+    "ウゥ/ u:",
+    "エェ/ e e",
+    "オォ/ o:",
+    "カァ/ k a:",
+    "キィ/ k i:",
+    "クゥ/ k u:",
+    "クャ/ ky a",
+    "クュ/ ky u",
+    "クョ/ ky o",
+    "ケェ/ k e:",
+    "コォ/ k o:",
+    "ガァ/ g a:",
+    "ギィ/ g i:",
+    "グゥ/ g u:",
+    "グャ/ gy a",
+    "グュ/ gy u",
+    "グョ/ gy o",
+    "ゲェ/ g e:",
+    "ゴォ/ g o:",
+    "サァ/ s a:",
+    "シィ/ sh i:",
+    "スゥ/ s u:",
+    "スャ/ sh a",
+    "スュ/ sh u",
+    "スョ/ sh o",
+    "セェ/ s e:",
+    "ソォ/ s o:",
+    "ザァ/ z a:",
+    "ジィ/ j i:",
+    "ズゥ/ z u:",
+    "ズャ/ zy a",
+    "ズュ/ zy u",
+    "ズョ/ zy o",
+    "ゼェ/ z e:",
+    "ゾォ/ z o:",
+    "タァ/ t a:",
+    "チィ/ ch i:",
+    "ツァ/ ts a",
+    "ツィ/ ts i",
+    "ツゥ/ ts u:",
+    "ツャ/ ch a",
+    "ツュ/ ch u",
+    "ツョ/ ch o",
+    "ツェ/ ts e",
+    "ツォ/ ts o",
+    "テェ/ t e:",
+    "トォ/ t o:",
+    "ダァ/ d a:",
+    "ヂィ/ j i:",
+    "ヅゥ/ d u:",
+    "ヅャ/ zy a",
+    "ヅュ/ zy u",
+    "ヅョ/ zy o",
+    "デェ/ d e:",
+    "ドォ/ d o:",
+    "ナァ/ n a:",
+    "ニィ/ n i:",
+    "ヌゥ/ n u:",
+    "ヌャ/ ny a",
+    "ヌュ/ ny u",
+    "ヌョ/ ny o",
+    "ネェ/ n e:",
+    "ノォ/ n o:",
+    "ハァ/ h a:",
+    "ヒィ/ h i:",
+    "フゥ/ f u:",
+    "フャ/ hy a",
+    "フュ/ hy u",
+    "フョ/ hy o",
+    "ヘェ/ h e:",
+    "ホォ/ h o:",
+    "バァ/ b a:",
+    "ビィ/ b i:",
+    "ブゥ/ b u:",
+    "フャ/ hy a",
+    "ブュ/ by u",
+    "フョ/ hy o",
+    "ベェ/ b e:",
+    "ボォ/ b o:",
+    "パァ/ p a:",
+    "ピィ/ p i:",
+    "プゥ/ p u:",
+    "プャ/ py a",
+    "プュ/ py u",
+    "プョ/ py o",
+    "ペェ/ p e:",
+    "ポォ/ p o:",
+    "マァ/ m a:",
+    "ミィ/ m i:",
+    "ムゥ/ m u:",
+    "ムャ/ my a",
+    "ムュ/ my u",
+    "ムョ/ my o",
+    "メェ/ m e:",
+    "モォ/ m o:",
+    "ヤァ/ y a:",
+    "ユゥ/ y u:",
+    "ユャ/ y a:",
+    "ユュ/ y u:",
+    "ユョ/ y o:",
+    "ヨォ/ y o:",
+    "ラァ/ r a:",
+    "リィ/ r i:",
+    "ルゥ/ r u:",
+    "ルャ/ ry a",
+    "��ュ/ ry u",
+    "ルョ/ ry o",
+    "レェ/ r e:",
+    "ロォ/ r o:",
+    "ワァ/ w a:",
+    "ヲォ/ o:",
+    "ディ/ d i",
+    "デェ/ d e:",
+    "デャ/ dy a",
+    "デュ/ dy u",
+    "デョ/ dy o",
+    "ティ/ t i",
+    "テェ/ t e:",
+    "テャ/ ty a",
+    "テュ/ ty u",
+    "テョ/ ty o",
+    "スィ/ s i",
+    "ズァ/ z u a",
+    "ズィ/ z i",
+    "ズゥ/ z u",
+    "ズャ/ zy a",
+    "ズュ/ zy u",
+    "ズョ/ zy o",
+    "ズェ/ z e",
+    "ズォ/ z o",
+    "キャ/ ky a",
+    "キュ/ ky u",
+    "キョ/ ky o",
+    "シャ/ sh a",
+    "シュ/ sh u",
+    "シェ/ sh e",
+    "ショ/ sh o",
+    "チャ/ ch a",
+    "チュ/ ch u",
+    "チェ/ ch e",
+    "チョ/ ch o",
+    "トゥ/ t u",
+    "トャ/ ty a",
+    "トュ/ ty u",
+    "トョ/ ty o",
+    "ドァ/ d o a",
+    "ドゥ/ d u",
+    "ドャ/ dy a",
+    "ドュ/ dy u",
+    "ドョ/ dy o",
+    "ドォ/ d o:",
+    "ニャ/ ny a",
+    "ニュ/ ny u",
+    "ニョ/ ny o",
+    "ヒャ/ hy a",
+    "ヒュ/ hy u",
+    "ヒョ/ hy o",
+    "ミャ/ my a",
+    "ミュ/ my u",
+    "ミョ/ my o",
+    "リャ/ ry a",
+    "リュ/ ry u",
+    "リョ/ ry o",
+    "ギャ/ gy a",
+    "ギュ/ gy u",
+    "ギョ/ gy o",
+    "ヂェ/ j e",
+    "ヂャ/ j a",
+    "ヂュ/ j u",
+    "ヂョ/ j o",
+    "ジェ/ j e",
+    "ジャ/ j a",
+    "ジュ/ j u",
+    "ジョ/ j o",
+    "ビャ/ by a",
+    "ビュ/ by u",
+    "ビョ/ by o",
+    "ピャ/ py a",
+    "ピュ/ py u",
+    "ピョ/ py o",
+    "ウァ/ u a",
+    "ウィ/ w i",
+    "ウェ/ w e",
+    "ウォ/ w o",
+    "ファ/ f a",
+    "フィ/ f i",
+    "フゥ/ f u",
+    "フャ/ hy a",
+    "フュ/ hy u",
+    "フョ/ hy o",
+    "フェ/ f e",
+    "フォ/ f o",
+    "ヴァ/ b a",
+    "ヴィ/ b i",
+    "ヴェ/ b e",
+    "ヴォ/ b o",
+    "ヴュ/ by u",
+    "アー/ a:",
+    "イー/ i:",
+    "ウー/ u:",
+    "エー/ e:",
+    "オー/ o:",
+    "カー/ k a:",
+    "キー/ k i:",
+    "クー/ k u:",
+    "ケー/ k e:",
+    "コー/ k o:",
+    "サー/ s a:",
+    "シー/ sh i:",
+    "スー/ s u:",
+    "セー/ s e:",
+    "ソー/ s o:",
+    "ター/ t a:",
+    "チー/ ch i:",
+    "ツー/ ts u:",
+    "テー/ t e:",
+    "トー/ t o:",
+    "ナー/ n a:",
+    "ニー/ n i:",
+    "ヌー/ n u:",
+    "ネー/ n e:",
+    "ノー/ n o:",
+    "ハー/ h a:",
+    "ヒー/ h i:",
+    "フー/ f u:",
+    "ヘー/ h e:",
+    "ホー/ h o:",
+    "マー/ m a:",
+    "ミー/ m i:",
+    "ムー/ m u:",
+    "メー/ m e:",
+    "モー/ m o:",
+    "ラー/ r a:",
+    "リー/ r i:",
+    "ルー/ r u:",
+    "レー/ r e:",
+    "ロー/ r o:",
+    "ガー/ g a:",
+    "ギー/ g i:",
+    "グー/ g u:",
+    "ゲー/ g e:",
+    "ゴー/ g o:",
+    "ザー/ z a:",
+    "ジー/ j i:",
+    "ズー/ z u:",
+    "ゼー/ z e:",
+    "ゾー/ z o:",
+    "ダー/ d a:",
+    "ヂー/ j i:",
+    "ヅー/ z u:",
+    "デー/ d e:",
+    "ドー/ d o:",
+    "バー/ b a:",
+    "ビー/ b i:",
+    "ブー/ b u:",
+    "ベー/ b e:",
+    "ボー/ b o:",
+    "パー/ p a:",
+    "ピー/ p i:",
+    "プー/ p u:",
+    "ペー/ p e:",
+    "ポー/ p o:",
+    "ヤー/ y a:",
+    "ユー/ y u:",
+    "ヨー/ y o:",
+    "ワー/ w a:",
+    "ヰー/ i:",
+    "ヱー/ e:",
+    "ヲー/ o:",
+    "ヴー/ b u:",
+    # Conversion of 1 letter
+    "ア/ a",
+    "イ/ i",
+    "ウ/ u",
+    "エ/ e",
+    "オ/ o",
+    "カ/ k a",
+    "キ/ k i",
+    "ク/ k u",
+    "ケ/ k e",
+    "コ/ k o",
+    "サ/ s a",
+    "シ/ sh i",
+    "ス/ s u",
+    "セ/ s e",
+    "ソ/ s o",
+    "タ/ t a",
+    "チ/ ch i",
+    "ツ/ ts u",
+    "テ/ t e",
+    "ト/ t o",
+    "ナ/ n a",
+    "ニ/ n i",
+    "ヌ/ n u",
+    "ネ/ n e",
+    "ノ/ n o",
+    "ハ/ h a",
+    "ヒ/ h i",
+    "フ/ f u",
+    "ヘ/ h e",
+    "ホ/ h o",
+    "マ/ m a",
+    "ミ/ m i",
+    "ム/ m u",
+    "メ/ m e",
+    "モ/ m o",
+    "ラ/ r a",
+    "リ/ r i",
+    "ル/ r u",
+    "レ/ r e",
+    "ロ/ r o",
+    "ガ/ g a",
+    "ギ/ g i",
+    "グ/ g u",
+    "ゲ/ g e",
+    "ゴ/ g o",
+    "ザ/ z a",
+    "ジ/ j i",
+    "ズ/ z u",
+    "ゼ/ z e",
+    "ゾ/ z o",
+    "ダ/ d a",
+    "ヂ/ j i",
+    "ヅ/ z u",
+    "デ/ d e",
+    "ド/ d o",
+    "バ/ b a",
+    "ビ/ b i",
+    "ブ/ b u",
+    "ベ/ b e",
+    "ボ/ b o",
+    "パ/ p a",
+    "ピ/ p i",
+    "プ/ p u",
+    "ペ/ p e",
+    "ポ/ p o",
+    "ヤ/ y a",
+    "ユ/ y u",
+    "ヨ/ y o",
+    "ワ/ w a",
+    "ヰ/ i",
+    "ヱ/ e",
+    "ヲ/ o",
+    "ン/ N",
+    "ッ/ q",
+    "ヴ/ b u",
+    "ー/:", #这个不起作用
+    # Try converting broken text
+    "ァ/ a",
+    "ィ/ i",
+    "ゥ/ u",
+    "ェ/ e",
+    "ォ/ o",
+    "ヮ/ w a",
+    "ォ/ o",
+    # Symbols
+    "、/ ,",
+    "。/ .",
+    "！/ !",
+    "？/ ?",
+    "・/ ,",
+]
+
+_COLON_RX = re.compile(":+")
+_REJECT_RX = re.compile("[^ a-zA-Z:,.?]")
+
+
+def _makerulemap():
+    l = [tuple(x.split("/")) for x in _CONVRULES]
+    return tuple({k: v for k, v in l if len(k) == i} for i in (1, 2))
+
+
+_RULEMAP1, _RULEMAP2 = _makerulemap()
+
+
+def kata2phoneme(text: str) -> str:
+    """Convert katakana text to phonemes."""
+    text = text.strip()
+    res = []
+    while text:
+        if len(text) >= 2:
+            x = _RULEMAP2.get(text[:2])
+            if x is not None:
+                text = text[2:]
+                res += x.split(" ")[1:]
+                continue
+        x = _RULEMAP1.get(text[0])
+        if x is not None:
+            text = text[1:]
+            res += x.split(" ")[1:]
+            continue
+        res.append(text[0])
+        text = text[1:]
+    # res = _COLON_RX.sub(":", res)
+    return res
+
+
+_KATAKANA = "".join(chr(ch) for ch in range(ord("ァ"), ord("ン") + 1))
+_HIRAGANA = "".join(chr(ch) for ch in range(ord("ぁ"), ord("ん") + 1))
+_HIRA2KATATRANS = str.maketrans(_HIRAGANA, _KATAKANA)
+
+
+def hira2kata(text: str) -> str:
+    text = text.translate(_HIRA2KATATRANS)
+    return text.replace("う゛", "ヴ")
+
+
+_SYMBOL_TOKENS = set(list("・、。？！"))
+_NO_YOMI_TOKENS = set(list("「」『』―（）［］[]"))
+_TAGGER = MeCab.Tagger()
+
+
+def text2kata(text: str) -> str:
+    parsed = _TAGGER.parse(text)
+    res = []
+    for line in parsed.split("\n"):
+        if line == "EOS":
+            break
+        parts = line.split("\t")
+
+        word, yomi = parts[0], parts[1]
+        if yomi:
+            res.append(yomi)
+        else:
+            if word in _SYMBOL_TOKENS:
+                res.append(word)
+            elif word in ("っ", "ッ"):
+                res.append("ッ")
+            elif word in _NO_YOMI_TOKENS:
+                pass
+            else:
+                res.append(word)
+    return hira2kata("".join(res))
+
+
+def text2sep_kata(text: str) -> (list, list):
+    parsed = _TAGGER.parse(text)
+    res = []
+    sep = []
+    for line in parsed.split("\n"):
+        if line == "EOS":
+            break
+        parts = line.split("\t")
+
+        word, yomi = parts[0], parts[1]
+        if yomi:
+            res.append(yomi)
+        else:
+            if word in _SYMBOL_TOKENS:
+                res.append(word)
+            elif word in ("っ", "ッ"):
+                res.append("ッ")
+            elif word in _NO_YOMI_TOKENS:
+                pass
+            else:
+                res.append(word)
+        sep.append(word)
+    return sep, [hira2kata(i) for i in res]
+
+
+_ALPHASYMBOL_YOMI = {
+    "#": "シャープ",
+    "%": "パーセント",
+    "&": "アンド",
+    "+": "プラス",
+    "-": "マイナス",
+    ":": "コロン",
+    ";": "セミコロン",
+    "<": "小なり",
+    "=": "イコール",
+    ">": "大なり",
+    "@": "アット",
+    "a": "エー",
+    "b": "ビー",
+    "c": "シー",
+    "d": "ディー",
+    "e": "イー",
+    "f": "エフ",
+    "g": "ジー",
+    "h": "エイチ",
+    "i": "アイ",
+    "j": "ジェー",
+    "k": "ケー",
+    "l": "エル",
+    "m": "エム",
+    "n": "エヌ",
+    "o": "オー",
+    "p": "ピー",
+    "q": "キュー",
+    "r": "アール",
+    "s": "エス",
+    "t": "ティー",
+    "u": "ユー",
+    "v": "ブイ",
+    "w": "ダブリュー",
+    "x": "エックス",
+    "y": "ワイ",
+    "z": "ゼット",
+    "α": "アルファ",
+    "β": "ベータ",
+    "γ": "ガンマ",
+    "δ": "デルタ",
+    "ε": "イプシロン",
+    "ζ": "ゼータ",
+    "η": "イータ",
+    "θ": "シータ",
+    "ι": "イオタ",
+    "κ": "カッパ",
+    "λ": "ラムダ",
+    "μ": "ミュー",
+    "ν": "ニュー",
+    "ξ": "クサイ",
+    "ο": "オミクロン",
+    "π": "パイ",
+    "ρ": "ロー",
+    "σ": "シグマ",
+    "τ": "タウ",
+    "υ": "ウプシロン",
+    "φ": "ファイ",
+    "χ": "カイ",
+    "ψ": "プサイ",
+    "ω": "オメガ",
+}
+
+
+_NUMBER_WITH_SEPARATOR_RX = re.compile("[0-9]{1,3}(,[0-9]{3})+")
+_CURRENCY_MAP = {"$": "ドル", "¥": "円", "£": "ポンド", "€": "ユーロ"}
+_CURRENCY_RX = re.compile(r"([$¥£€])([0-9.]*[0-9])")
+_NUMBER_RX = re.compile(r"[0-9]+(\.[0-9]+)?")
+
+
+def japanese_convert_numbers_to_words(text: str) -> str:
+    res = _NUMBER_WITH_SEPARATOR_RX.sub(lambda m: m[0].replace(",", ""), text)
+    res = _CURRENCY_RX.sub(lambda m: m[2] + _CURRENCY_MAP.get(m[1], m[1]), res)
+    res = _NUMBER_RX.sub(lambda m: num2words(m[0], lang="ja"), res)
+    return res
+
+
+def japanese_convert_alpha_symbols_to_words(text: str) -> str:
+    return "".join([_ALPHASYMBOL_YOMI.get(ch, ch) for ch in text.lower()])
+
+
+def japanese_text_to_phonemes(text: str) -> str:
+    """Convert Japanese text to phonemes."""
+    res = unicodedata.normalize("NFKC", text)
+    res = japanese_convert_numbers_to_words(res)
+    # res = japanese_convert_alpha_symbols_to_words(res)
+    res = text2kata(res)
+    res = kata2phoneme(res)
+    return res
+
+
+def is_japanese_character(char):
+    # 定义日语文字系统的 Unicode 范围
+    japanese_ranges = [
+        (0x3040, 0x309F),  # 平假名
+        (0x30A0, 0x30FF),  # 片假名
+        (0x4E00, 0x9FFF),  # 汉字 (CJK Unified Ideographs)
+        (0x3400, 0x4DBF),  # 汉字扩展 A
+        (0x20000, 0x2A6DF),  # 汉字扩展 B
+        # 可以根据需要添加其他汉字扩展范围
+    ]
+
+    # 将字符的 Unicode 编码转换为整数
+    char_code = ord(char)
+
+    # 检查字符是否在任何一个日语范围内
+    for start, end in japanese_ranges:
+        if start <= char_code <= end:
+            return True
+
+    return False
+
+
+rep_map = {
+    "：": ",",
+    "；": ",",
+    "，": ",",
+    "。": ".",
+    "！": "!",
+    "？": "?",
+    "\n": ".",
+    "·": ",",
+    "、": ",",
+    "…": "...",
+}
+
+
+def replace_punctuation(text):
+    pattern = re.compile("|".join(re.escape(p) for p in rep_map.keys()))
+
+    replaced_text = pattern.sub(lambda x: rep_map[x.group()], text)
+
+    replaced_text = re.sub(
+        r"[^\u3040-\u309F\u30A0-\u30FF\u4E00-\u9FFF\u3400-\u4DBF"
+        + "".join(punctuation)
+        + r"]+",
+        "",
+        replaced_text,
+    )
+
+    return replaced_text
+
+
+def text_normalize(text):
+    res = unicodedata.normalize("NFKC", text)
+    res = japanese_convert_numbers_to_words(res)
+    # res = "".join([i for i in res if is_japanese_character(i)])
+    res = replace_punctuation(res)
+    return res
+
+
+def distribute_phone(n_phone, n_word):
+    phones_per_word = [0] * n_word
+    for task in range(n_phone):
+        min_tasks = min(phones_per_word)
+        min_index = phones_per_word.index(min_tasks)
+        phones_per_word[min_index] += 1
+    return phones_per_word
+
+
+tokenizer = AutoTokenizer.from_pretrained("./bert/bert-base-japanese-v3")
+
+
+def g2p(norm_text):
+    sep_text, sep_kata = text2sep_kata(norm_text)
+    sep_tokenized = [tokenizer.tokenize(i) for i in sep_text]
+    sep_phonemes = [kata2phoneme(i) for i in sep_kata]
+    # 异常处理，MeCab不认识的词的话会一路传到这里来，然后炸掉。目前来看只有那些超级稀有的生僻词会出现这种情况
+    for i in sep_phonemes:
+        for j in i:
+            assert j in symbols, (sep_text, sep_kata, sep_phonemes)
+
+    word2ph = []
+    for token, phoneme in zip(sep_tokenized, sep_phonemes):
+        phone_len = len(phoneme)
+        word_len = len(token)
+
+        aaa = distribute_phone(phone_len, word_len)
+        word2ph += aaa
+    phones = ["_"] + [j for i in sep_phonemes for j in i] + ["_"]
+    tones = [0 for i in phones]
+    word2ph = [1] + word2ph + [1]
+    return phones, tones, word2ph
+
+if __name__ == "__main__":
+    tokenizer = AutoTokenizer.from_pretrained("./bert/bert-base-japanese-v3")
+    text = "だったら私、スズカさんと同じチームに入りたいです！ スズカさんの走りを毎日近くで、なんなら真横から見ていたいので！"
+    #print(_TAGGER.parse(text))
+   # nodes = [{"surface": "こんにちは", "pos": "感動詞:*:*:*", "pron": "コンニチワ", "c_type": "*", "c_form": "*", "accent_type": 0, "accent_con_type": "-1", "chain_flag": -1}]
+    nodes = [{"surface":"こんにちは","pron": "コンニチワ","pos": "感動詞:*:*:*",}]
+    from text.japanese_bert import get_bert_feature
+    import pyopenjtalk
+    from marine.predict import Predictor
+    from marine.utils.openjtalk_util import convert_njd_feature_to_marine_feature
+    text = text_normalize(text)
+    NJD_NODES = pyopenjtalk.run_frontend(text)
+    predictor = Predictor()
+ #   important_info = [{"string":i["string"],"pron":i["pron"],"acc":i["acc"]}for i in pyopenjtalk.estimate_accent(NJD_NODES)]
+    print(text)
+    
+    marine_feature = convert_njd_feature_to_marine_feature(NJD_NODES)
+    results = predictor.predict([marine_feature])
+    for mora,acc in zip(results["mora"][0],results["accent_status"][0]):
+        print(f"{mora}:{acc}")
+    # for i in pyopenjtalk.estimate_accent(NJD_NODES):
+    #     print(f"{i['string']}:{i['pron']}:{i['acc']}")
+#     info = pyopenjtalk.extract_fullcontext(text,run_marine=True)
+#     info_nomarine = pyopenjtalk.extract_fullcontext(text,run_marine=False)
+#    # nodes = pyopenjtalk
+#    # print(info)
+#     for i,j in zip(info,info_nomarine):
+#         print(i)
+#         print(j)
+#         print("\n")
+    # predictor = Predictor()
+    #print(pyopenjtalk.estimate_accent(text))
+    # output = predictor.predict([nodes],accent_represent_mode="high_low")
+    #print(output)
+    # phones, tones, word2ph = g2p(text)
+    # bert = get_bert_feature(text, word2ph)
+
+    # print(phones, tones, word2ph, bert.shape)

text/japanese_bert.py

@@ -1,68 +1,87 @@
-import sys
-
-import torch
-from transformers import AutoModelForMaskedLM, AutoTokenizer
-
-from config import config
-from text.japanese import text2sep_kata
-
-LOCAL_PATH = "./bert/deberta-v2-large-japanese-char-wwm"
-
-tokenizer = AutoTokenizer.from_pretrained(LOCAL_PATH)
-
-models = dict()
-
-
-def get_bert_feature(
-    text,
-    word2ph,
-    device=config.bert_gen_config.device,
-    style_text=None,
-    style_weight=0.7,
-):
-    text = "".join(text2sep_kata(text)[0])
-    if style_text:
-        style_text = "".join(text2sep_kata(style_text)[0])
-    if (
-        sys.platform == "darwin"
-        and torch.backends.mps.is_available()
-        and device == "cpu"
-    ):
-        device = "mps"
-    if not device:
-        device = "cuda"
-    if device not in models.keys():
-        if config.webui_config.fp16_run:
-            models[device] = AutoModelForMaskedLM.from_pretrained(LOCAL_PATH, torch_dtype=torch.float16).to(device)
-        else:
-            models[device] = AutoModelForMaskedLM.from_pretrained(LOCAL_PATH).to(device)
-    with torch.no_grad():
-        inputs = tokenizer(text, return_tensors="pt")
-        for i in inputs:
-            inputs[i] = inputs[i].to(device)
-        res = models[device](**inputs, output_hidden_states=True)
-        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].float().cpu()
-        if style_text:
-            style_inputs = tokenizer(style_text, return_tensors="pt")
-            for i in style_inputs:
-                style_inputs[i] = style_inputs[i].to(device)
-            style_res = models[device](**style_inputs, output_hidden_states=True)
-            style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].float().cpu()
-            style_res_mean = style_res.mean(0)
-
-    assert len(word2ph) == len(text) + 2
-    word2phone = word2ph
-    phone_level_feature = []
-    for i in range(len(word2phone)):
-        if style_text:
-            repeat_feature = (
-                res[i].repeat(word2phone[i], 1) * (1 - style_weight)
-                + style_res_mean.repeat(word2phone[i], 1) * style_weight
-            )
-        else:
-            repeat_feature = res[i].repeat(word2phone[i], 1)
-        phone_level_feature.append(repeat_feature)
-
-    phone_level_feature = torch.cat(phone_level_feature, dim=0)
-
-    return phone_level_feature.T
+import torch
+from transformers import AutoTokenizer, AutoModelForMaskedLM
+import sys
+import os
+from text.japanese import text2sep_kata
+tokenizer = AutoTokenizer.from_pretrained("./bert/bert-base-japanese-v3")
+
+models = dict()
+
+
+def get_bert_feature(text, word2ph, device=None):
+    sep_text,_ = text2sep_kata(text)
+    sep_tokens = [tokenizer.tokenize(t) for t in sep_text]
+    sep_ids = [tokenizer.convert_tokens_to_ids(t) for t in sep_tokens]
+    sep_ids = [2]+[item for sublist in sep_ids for item in sublist]+[3]
+    return get_bert_feature_with_token(sep_ids, word2ph, device)
+
+
+# def get_bert_feature(text, word2ph, device=None):
+#     if (
+#         sys.platform == "darwin"
+#         and torch.backends.mps.is_available()
+#         and device == "cpu"
+#     ):
+#         device = "mps"
+#     if not device:
+#         device = "cuda"
+#     if device not in models.keys():
+#         models[device] = AutoModelForMaskedLM.from_pretrained(
+#             "cl-tohoku/bert-base-japanese-v3"
+#         ).to(device)
+#     with torch.no_grad():
+#         inputs = tokenizer(text, return_tensors="pt")
+#         for i in inputs:
+#             inputs[i] = inputs[i].to(device)
+#         res = models[device](**inputs, output_hidden_states=True)
+#         res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
+#     assert inputs["input_ids"].shape[-1] == len(word2ph)
+#     word2phone = word2ph
+#     phone_level_feature = []
+#     for i in range(len(word2phone)):
+#         repeat_feature = res[i].repeat(word2phone[i], 1)
+#         phone_level_feature.append(repeat_feature)
+
+#     phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+#     return phone_level_feature.T
+
+def get_bert_feature_with_token(tokens, word2ph, device=None):
+    if (
+        sys.platform == "darwin"
+        and torch.backends.mps.is_available()
+        and device == "cpu"
+    ):
+        device = "mps"
+    if not device:
+        device = "cuda"
+    if device not in models.keys():
+        models[device] = AutoModelForMaskedLM.from_pretrained(
+            "./bert/bert-base-japanese-v3"
+        ).to(device)
+    with torch.no_grad():
+        inputs = torch.tensor(tokens).to(device).unsqueeze(0)
+        token_type_ids = torch.zeros_like(inputs).to(device)
+        attention_mask = torch.ones_like(inputs).to(device)
+        inputs = {"input_ids": inputs, "token_type_ids": token_type_ids, "attention_mask": attention_mask}
+
+
+        # for i in inputs:
+        #     inputs[i] = inputs[i].to(device)
+        res = models[device](**inputs, output_hidden_states=True)
+        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
+    assert inputs["input_ids"].shape[-1] == len(word2ph)
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+    return phone_level_feature.T
+
+
+if __name__ == "__main__":
+    print(get_bert_feature("観覧車",[4,2]))
+    pass