Refactor: improve regular expression performance during NLP

Improve performance by pre-compiling regular expressions that are executed many times.

style_bert_vits2/nlp/chinese/normalizer.py

@@ -5,17 +5,7 @@ import cn2an
 from style_bert_vits2.nlp.symbols import PUNCTUATIONS
 
 
-def normalize_text(text: str) -> str:
+__REPLACE_MAP = {
-    numbers = re.findall(r"\d+(?:\.?\d+)?", text)
-    for number in numbers:
-        text = text.replace(number, cn2an.an2cn(number), 1)
-    text = replace_punctuation(text)
-    return text
-
-
-def replace_punctuation(text: str) -> str:
-
-    REPLACE_MAP = {
     "：": ",",
     "；": ",",
     "，": ",",
@@ -49,10 +39,21 @@ def replace_punctuation(text: str) -> str:
     "」": "'",
 }
 
-    text = text.replace("嗯", "恩").replace("呣", "母")
-    pattern = re.compile("|".join(re.escape(p) for p in REPLACE_MAP.keys()))
 
-    replaced_text = pattern.sub(lambda x: REPLACE_MAP[x.group()], text)
+def normalize_text(text: str) -> str:
+    numbers = re.findall(r"\d+(?:\.?\d+)?", text)
+    for number in numbers:
+        text = text.replace(number, cn2an.an2cn(number), 1)
+    text = replace_punctuation(text)
+    return text
+
+
+def replace_punctuation(text: str) -> str:
+
+    text = text.replace("嗯", "恩").replace("呣", "母")
+    pattern = re.compile("|".join(re.escape(p) for p in __REPLACE_MAP.keys()))
+
+    replaced_text = pattern.sub(lambda x: __REPLACE_MAP[x.group()], text)
 
     replaced_text = re.sub(
         r"[^\u4e00-\u9fa5" + "".join(PUNCTUATIONS) + r"]+", "", replaced_text

style_bert_vits2/nlp/japanese/g2p.py

@@ -5,7 +5,7 @@ from style_bert_vits2.constants import Languages
 from style_bert_vits2.logging import logger
 from style_bert_vits2.nlp import bert_models
 from style_bert_vits2.nlp.japanese import pyopenjtalk_worker as pyopenjtalk
-from style_bert_vits2.nlp.japanese.mora_list import MORA_KATA_TO_MORA_PHONEMES
+from style_bert_vits2.nlp.japanese.mora_list import MORA_KATA_TO_MORA_PHONEMES, VOWELS
 from style_bert_vits2.nlp.japanese.normalizer import replace_punctuation
 from style_bert_vits2.nlp.symbols import PUNCTUATIONS
 
@@ -144,7 +144,7 @@ def text_to_sep_kata(
                 ## 例外を送出しない場合
                 ## 読めない文字は「'」として扱う
                 logger.warning(
-                    f"Cannot read: {word} in:\n{norm_text}, replaced with \"'\""
+                    f'Cannot read: {word} in:\n{norm_text}, replaced with "\'"'
                 )
                 # word の文字数分「'」を追加
                 yomi = "'" * len(word)
@@ -428,15 +428,23 @@ def __g2phone_tone_wo_punct(text: str) -> list[tuple[str, int]]:
     return result
 
 
+__PYOPENJTALK_G2P_PROSODY_A1_PATTERN = re.compile(r"/A:([0-9\-]+)\+")
+__PYOPENJTALK_G2P_PROSODY_A2_PATTERN = re.compile(r"\+(\d+)\+")
+__PYOPENJTALK_G2P_PROSODY_A3_PATTERN = re.compile(r"\+(\d+)/")
+__PYOPENJTALK_G2P_PROSODY_E3_PATTERN = re.compile(r"!(\d+)_")
+__PYOPENJTALK_G2P_PROSODY_F1_PATTERN = re.compile(r"/F:(\d+)_")
+__PYOPENJTALK_G2P_PROSODY_P3_PATTERN = re.compile(r"\-(.*?)\+")
+
+
 def __pyopenjtalk_g2p_prosody(
     text: str, drop_unvoiced_vowels: bool = True
 ) -> list[str]:
     """
-    ESPnet の実装から引用、変更点無し。「ん」は「N」なことに注意。
+    ESPnet の実装から引用、概ね変更点無し。「ん」は「N」なことに注意。
     ref: https://github.com/espnet/espnet/blob/master/espnet2/text/phoneme_tokenizer.py
     ------------------------------------------------------------------------------------------
 
-    Extract phoneme + prosoody symbol sequence from input full-context labels.
+    Extract phoneme + prosody symbol sequence from input full-context labels.
 
     The algorithm is based on `Prosodic features control by symbols as input of
     sequence-to-sequence acoustic modeling for neural TTS`_ with some r9y9's tweaks.
@@ -457,8 +465,8 @@ def __pyopenjtalk_g2p_prosody(
         modeling for neural TTS`: https://doi.org/10.1587/transinf.2020EDP7104
     """
 
-    def _numeric_feature_by_regex(regex: str, s: str) -> int:
+    def _numeric_feature_by_regex(pattern: re.Pattern[str], s: str) -> int:
-        match = re.search(regex, s)
+        match = pattern.search(s)
         if match is None:
             return -50
         return int(match.group(1))
@@ -471,7 +479,7 @@ def __pyopenjtalk_g2p_prosody(
         lab_curr = labels[n]
 
         # current phoneme
-        p3 = re.search(r"\-(.*?)\+", lab_curr).group(1)  # type: ignore
+        p3 = __PYOPENJTALK_G2P_PROSODY_P3_PATTERN.search(lab_curr).group(1)  # type: ignore
         # deal unvoiced vowels as normal vowels
         if drop_unvoiced_vowels and p3 in "AEIOU":
             p3 = p3.lower()
@@ -483,7 +491,9 @@ def __pyopenjtalk_g2p_prosody(
                 phones.append("^")
             elif n == N - 1:
                 # check question form or not
-                e3 = _numeric_feature_by_regex(r"!(\d+)_", lab_curr)
+                e3 = _numeric_feature_by_regex(
+                    __PYOPENJTALK_G2P_PROSODY_E3_PATTERN, lab_curr
+                )
                 if e3 == 0:
                     phones.append("$")
                 elif e3 == 1:
@@ -496,14 +506,16 @@ def __pyopenjtalk_g2p_prosody(
             phones.append(p3)
 
         # accent type and position info (forward or backward)
-        a1 = _numeric_feature_by_regex(r"/A:([0-9\-]+)\+", lab_curr)
+        a1 = _numeric_feature_by_regex(__PYOPENJTALK_G2P_PROSODY_A1_PATTERN, lab_curr)
-        a2 = _numeric_feature_by_regex(r"\+(\d+)\+", lab_curr)
+        a2 = _numeric_feature_by_regex(__PYOPENJTALK_G2P_PROSODY_A2_PATTERN, lab_curr)
-        a3 = _numeric_feature_by_regex(r"\+(\d+)/", lab_curr)
+        a3 = _numeric_feature_by_regex(__PYOPENJTALK_G2P_PROSODY_A3_PATTERN, lab_curr)
 
         # number of mora in accent phrase
-        f1 = _numeric_feature_by_regex(r"/F:(\d+)_", lab_curr)
+        f1 = _numeric_feature_by_regex(__PYOPENJTALK_G2P_PROSODY_F1_PATTERN, lab_curr)
 
-        a2_next = _numeric_feature_by_regex(r"\+(\d+)\+", labels[n + 1])
+        a2_next = _numeric_feature_by_regex(
+            __PYOPENJTALK_G2P_PROSODY_A2_PATTERN, labels[n + 1]
+        )
         # accent phrase border
         if a3 == 1 and a2_next == 1 and p3 in "aeiouAEIOUNcl":
             phones.append("#")
@@ -560,9 +572,6 @@ def __handle_long(sep_phonemes: list[list[str]]) -> list[list[str]]:
         list[list[str]]: 長音記号を処理した音素のリストのリスト
     """
 
-    # 母音の集合 (便宜上「ん」を含める)
-    VOWELS = {"a", "i", "u", "e", "o", "N"}
-
     for i in range(len(sep_phonemes)):
         if len(sep_phonemes[i]) == 0:
             # 空白文字等でリストが空の場合
@@ -588,6 +597,15 @@ def __handle_long(sep_phonemes: list[list[str]]) -> list[list[str]]:
     return sep_phonemes
 
 
+__KATAKANA_PATTERN = re.compile(r"[\u30A0-\u30FF]+")
+__MORA_PATTERN = re.compile(
+    "|".join(
+        map(re.escape, sorted(MORA_KATA_TO_MORA_PHONEMES.keys(), key=len, reverse=True))
+    )
+)
+__LONG_PATTERN = re.compile(r"(\w)(ー*)")
+
+
 def __kata_to_phoneme_list(text: str) -> list[str]:
     """
     原則カタカナの `text` を受け取り、それをそのままいじらずに音素記号のリストに変換。
@@ -610,23 +628,20 @@ def __kata_to_phoneme_list(text: str) -> list[str]:
     if set(text).issubset(set(PUNCTUATIONS)):
         return list(text)
     # `text` がカタカナ（`ー`含む）のみからなるかどうかをチェック
-    if re.fullmatch(r"[\u30A0-\u30FF]+", text) is None:
+    if __KATAKANA_PATTERN.fullmatch(text) is None:
         raise ValueError(f"Input must be katakana only: {text}")
-    sorted_keys = sorted(MORA_KATA_TO_MORA_PHONEMES.keys(), key=len, reverse=True)
-    pattern = "|".join(map(re.escape, sorted_keys))
 
     def mora2phonemes(mora: str) -> str:
-        cosonant, vowel = MORA_KATA_TO_MORA_PHONEMES[mora]
+        consonant, vowel = MORA_KATA_TO_MORA_PHONEMES[mora]
-        if cosonant is None:
+        if consonant is None:
             return f" {vowel}"
-        return f" {cosonant} {vowel}"
+        return f" {consonant} {vowel}"
 
-    spaced_phonemes = re.sub(pattern, lambda m: mora2phonemes(m.group()), text)
+    spaced_phonemes = __MORA_PATTERN.sub(lambda m: mora2phonemes(m.group()), text)
 
     # 長音記号「ー」の処理
-    long_pattern = r"(\w)(ー*)"
     long_replacement = lambda m: m.group(1) + (" " + m.group(1)) * len(m.group(2))  # type: ignore
-    spaced_phonemes = re.sub(long_pattern, long_replacement, spaced_phonemes)
+    spaced_phonemes = __LONG_PATTERN.sub(long_replacement, spaced_phonemes)
 
     return spaced_phonemes.strip().split(" ")
 

style_bert_vits2/nlp/japanese/g2p_utils.py

@@ -1,5 +1,6 @@
 from style_bert_vits2.nlp.japanese.g2p import g2p
 from style_bert_vits2.nlp.japanese.mora_list import (
+    CONSONANTS,
     MORA_KATA_TO_MORA_PHONEMES,
     MORA_PHONEMES_TO_MORA_KATA,
 )
@@ -33,15 +34,6 @@ def phone_tone2kata_tone(phone_tone: list[tuple[str, int]]) -> list[tuple[str, i
         カタカナと音高のリスト。
     """
 
-    # 子音の集合
-    CONSONANTS = set(
-        [
-            consonant
-            for consonant, _ in MORA_KATA_TO_MORA_PHONEMES.values()
-            if consonant is not None
-        ]
-    )
-
     phone_tone = phone_tone[1:]  # 最初の("_", 0)を無視
     phones = [phone for phone, _ in phone_tone]
     tones = [tone for _, tone in phone_tone]

style_bert_vits2/nlp/japanese/mora_list.py

@@ -234,3 +234,15 @@ MORA_KATA_TO_MORA_PHONEMES: dict[str, tuple[Optional[str], str]] = {
     kana: (consonant, vowel)
     for [kana, consonant, vowel] in __MORA_LIST_MINIMUM + __MORA_LIST_ADDITIONAL
 }
+
+# 子音の集合
+CONSONANTS = set(
+    [
+        consonant
+        for consonant, _ in MORA_KATA_TO_MORA_PHONEMES.values()
+        if consonant is not None
+    ]
+)
+
+# 母音の集合 (便宜上「ん」を含める)
+VOWELS = {"a", "i", "u", "e", "o", "N"}

style_bert_vits2/nlp/japanese/normalizer.py

@@ -6,6 +6,81 @@ from num2words import num2words
 from style_bert_vits2.nlp.symbols import PUNCTUATIONS
 
 
+# 記号類の正規化マップ
+__REPLACE_MAP = {
+    "：": ",",
+    "；": ",",
+    "，": ",",
+    "。": ".",
+    "！": "!",
+    "？": "?",
+    "\n": ".",
+    "．": ".",
+    "…": "...",
+    "···": "...",
+    "・・・": "...",
+    "·": ",",
+    "・": ",",
+    "、": ",",
+    "$": ".",
+    "“": "'",
+    "”": "'",
+    '"': "'",
+    "‘": "'",
+    "’": "'",
+    "（": "'",
+    "）": "'",
+    "(": "'",
+    ")": "'",
+    "《": "'",
+    "》": "'",
+    "【": "'",
+    "】": "'",
+    "[": "'",
+    "]": "'",
+    # NFKC 正規化後のハイフン・ダッシュの変種を全て通常半角ハイフン - \u002d に変換
+    "\u02d7": "\u002d",  # ˗, Modifier Letter Minus Sign
+    "\u2010": "\u002d",  # ‐, Hyphen,
+    # "\u2011": "\u002d",  # ‑, Non-Breaking Hyphen, NFKC により \u2010 に変換される
+    "\u2012": "\u002d",  # ‒, Figure Dash
+    "\u2013": "\u002d",  # –, En Dash
+    "\u2014": "\u002d",  # —, Em Dash
+    "\u2015": "\u002d",  # ―, Horizontal Bar
+    "\u2043": "\u002d",  # ⁃, Hyphen Bullet
+    "\u2212": "\u002d",  # −, Minus Sign
+    "\u23af": "\u002d",  # ⎯, Horizontal Line Extension
+    "\u23e4": "\u002d",  # ⏤, Straightness
+    "\u2500": "\u002d",  # ─, Box Drawings Light Horizontal
+    "\u2501": "\u002d",  # ━, Box Drawings Heavy Horizontal
+    "\u2e3a": "\u002d",  # ⸺, Two-Em Dash
+    "\u2e3b": "\u002d",  # ⸻, Three-Em Dash
+    # "～": "-",  # これは長音記号「ー」として扱うよう変更
+    # "~": "-",  # これも長音記号「ー」として扱うよう変更
+    "「": "'",
+    "」": "'",
+}
+# 記号類の正規化パターン
+__REPLACE_PATTERN = re.compile("|".join(re.escape(p) for p in __REPLACE_MAP.keys()))
+# 句読点等の正規化パターン
+__PUNCTUATION_CLEANUP_PATTERN = re.compile(
+    # ↓ ひらがな、カタカナ、漢字
+    r"[^\u3040-\u309F\u30A0-\u30FF\u4E00-\u9FFF\u3400-\u4DBF\u3005"
+    # ↓ 半角アルファベット（大文字と小文字）
+    + r"\u0041-\u005A\u0061-\u007A"
+    # ↓ 全角アルファベット（大文字と小文字）
+    + r"\uFF21-\uFF3A\uFF41-\uFF5A"
+    # ↓ ギリシャ文字
+    + r"\u0370-\u03FF\u1F00-\u1FFF"
+    # ↓ "!", "?", "…", ",", ".", "'", "-", 但し`…`はすでに`...`に変換されている
+    + "".join(PUNCTUATIONS) + r"]+",  # fmt: skip
+)
+# 数字・通貨記号の正規化パターン
+__CURRENCY_MAP = {"$": "ドル", "¥": "円", "£": "ポンド", "€": "ユーロ"}
+__CURRENCY_PATTERN = re.compile(r"([$¥£€])([0-9.]*[0-9])")
+__NUMBER_PATTERN = re.compile(r"[0-9]+(\.[0-9]+)?")
+__NUMBER_WITH_SEPARATOR_PATTERN = re.compile("[0-9]{1,3}(,[0-9]{3})+")
+
+
 def normalize_text(text: str) -> str:
     """
     日本語のテキストを正規化する。
@@ -62,80 +137,11 @@ def replace_punctuation(text: str) -> str:
         str: 正規化されたテキスト
     """
 
-    # 記号類の正規化変換マップ
-    REPLACE_MAP = {
-        "：": ",",
-        "；": ",",
-        "，": ",",
-        "。": ".",
-        "！": "!",
-        "？": "?",
-        "\n": ".",
-        "．": ".",
-        "…": "...",
-        "···": "...",
-        "・・・": "...",
-        "·": ",",
-        "・": ",",
-        "、": ",",
-        "$": ".",
-        "“": "'",
-        "”": "'",
-        '"': "'",
-        "‘": "'",
-        "’": "'",
-        "（": "'",
-        "）": "'",
-        "(": "'",
-        ")": "'",
-        "《": "'",
-        "》": "'",
-        "【": "'",
-        "】": "'",
-        "[": "'",
-        "]": "'",
-        # NFKC 正規化後のハイフン・ダッシュの変種を全て通常半角ハイフン - \u002d に変換
-        "\u02d7": "\u002d",  # ˗, Modifier Letter Minus Sign
-        "\u2010": "\u002d",  # ‐, Hyphen,
-        # "\u2011": "\u002d",  # ‑, Non-Breaking Hyphen, NFKC により \u2010 に変換される
-        "\u2012": "\u002d",  # ‒, Figure Dash
-        "\u2013": "\u002d",  # –, En Dash
-        "\u2014": "\u002d",  # —, Em Dash
-        "\u2015": "\u002d",  # ―, Horizontal Bar
-        "\u2043": "\u002d",  # ⁃, Hyphen Bullet
-        "\u2212": "\u002d",  # −, Minus Sign
-        "\u23af": "\u002d",  # ⎯, Horizontal Line Extension
-        "\u23e4": "\u002d",  # ⏤, Straightness
-        "\u2500": "\u002d",  # ─, Box Drawings Light Horizontal
-        "\u2501": "\u002d",  # ━, Box Drawings Heavy Horizontal
-        "\u2e3a": "\u002d",  # ⸺, Two-Em Dash
-        "\u2e3b": "\u002d",  # ⸻, Three-Em Dash
-        # "～": "-",  # これは長音記号「ー」として扱うよう変更
-        # "~": "-",  # これも長音記号「ー」として扱うよう変更
-        "「": "'",
-        "」": "'",
-    }
-
-    pattern = re.compile("|".join(re.escape(p) for p in REPLACE_MAP.keys()))
-
     # 句読点を辞書で置換
-    replaced_text = pattern.sub(lambda x: REPLACE_MAP[x.group()], text)
+    replaced_text = __REPLACE_PATTERN.sub(lambda x: __REPLACE_MAP[x.group()], text)
 
-    replaced_text = re.sub(
-        # ↓ ひらがな、カタカナ、漢字
-        r"[^\u3040-\u309F\u30A0-\u30FF\u4E00-\u9FFF\u3400-\u4DBF\u3005"
-        # ↓ 半角アルファベット（大文字と小文字）
-        + r"\u0041-\u005A\u0061-\u007A"
-        # ↓ 全角アルファベット（大文字と小文字）
-        + r"\uFF21-\uFF3A\uFF41-\uFF5A"
-        # ↓ ギリシャ文字
-        + r"\u0370-\u03FF\u1F00-\u1FFF"
-        # ↓ "!", "?", "…", ",", ".", "'", "-", 但し`…`はすでに`...`に変換されている
-        + "".join(PUNCTUATIONS) + r"]+",
     # 上述以外の文字を削除
-        "",
+    replaced_text = __PUNCTUATION_CLEANUP_PATTERN.sub("", replaced_text)
-        replaced_text,
-    )
 
     return replaced_text
 
@@ -151,13 +157,8 @@ def __convert_numbers_to_words(text: str) -> str:
         str: 変換されたテキスト
     """
 
-    NUMBER_WITH_SEPARATOR_PATTERN = re.compile("[0-9]{1,3}(,[0-9]{3})+")
+    res = __NUMBER_WITH_SEPARATOR_PATTERN.sub(lambda m: m[0].replace(",", ""), text)
-    CURRENCY_MAP = {"$": "ドル", "¥": "円", "£": "ポンド", "€": "ユーロ"}
+    res = __CURRENCY_PATTERN.sub(lambda m: m[2] + __CURRENCY_MAP.get(m[1], m[1]), res)
-    CURRENCY_PATTERN = re.compile(r"([$¥£€])([0-9.]*[0-9])")
+    res = __NUMBER_PATTERN.sub(lambda m: num2words(m[0], lang="ja"), res)
-    NUMBER_PATTERN = re.compile(r"[0-9]+(\.[0-9]+)?")
-
-    res = NUMBER_WITH_SEPARATOR_PATTERN.sub(lambda m: m[0].replace(",", ""), text)
-    res = CURRENCY_PATTERN.sub(lambda m: m[2] + CURRENCY_MAP.get(m[1], m[1]), res)
-    res = NUMBER_PATTERN.sub(lambda m: num2words(m[0], lang="ja"), res)
 
     return res