更新最新的中文多音字处理逻辑

更新最新的中文多音字处理逻辑

GPT_SoVITS/text/chinese2.py

@@ -4,6 +4,7 @@ import re
 
 import cn2an
 from pypinyin import lazy_pinyin, Style
+from pypinyin.contrib.tone_convert import to_normal, to_finals_tone3, to_initials, to_finals
 
 from text.symbols import punctuation
 from text.tone_sandhi import ToneSandhi
@@ -24,7 +25,7 @@ import jieba_fast.posseg as psg
 is_g2pw = True#True if is_g2pw_str.lower() == 'true' else False
 if is_g2pw:
     print("当前使用g2pw进行拼音推理")
-    from text.g2pw import G2PWPinyin
+    from text.g2pw import G2PWPinyin, correct_pronunciation
     parent_directory = os.path.dirname(current_file_path)
     g2pw = G2PWPinyin(model_dir="GPT_SoVITS/text/G2PWModel",model_source="GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large",v_to_u=False, neutral_tone_with_five=True)
 
@@ -42,6 +43,8 @@ rep_map = {
     "$": ".",
     "/": ",",
     "—": "-",
+    "~": "…",
+    "～":"…",
 }
 
 tone_modifier = ToneSandhi()
@@ -60,26 +63,6 @@ def replace_punctuation(text):
     return replaced_text
 
 
-def replace_punctuation_with_en(text):
-    text = text.replace("嗯", "恩").replace("呣", "母")
-    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"[^\u4e00-\u9fa5A-Za-z" + "".join(punctuation) + r"]+", "", replaced_text
-    )
-
-    return replaced_text
-
-
-def replace_consecutive_punctuation(text):
-    punctuations = ''.join(re.escape(p) for p in punctuation)
-    pattern = f'([{punctuations}])([{punctuations}])+'
-    result = re.sub(pattern, r'\1', text)
-    return result
-
-
 def g2p(text):
     pattern = r"(?<=[{0}])\s*".format("".join(punctuation))
     sentences = [i for i in re.split(pattern, text) if i.strip() != ""]
@@ -91,16 +74,10 @@ def _get_initials_finals(word):
     initials = []
     finals = []
 
-    if not is_g2pw:
+    orig_initials = lazy_pinyin(word, neutral_tone_with_five=True, style=Style.INITIALS)
-        orig_initials = lazy_pinyin(word, neutral_tone_with_five=True, style=Style.INITIALS)
+    orig_finals = lazy_pinyin(
-        orig_finals = lazy_pinyin(
+        word, neutral_tone_with_five=True, style=Style.FINALS_TONE3
-            word, neutral_tone_with_five=True, style=Style.FINALS_TONE3
+    )
-        )
-    else:
-        orig_initials = g2pw.lazy_pinyin(word, neutral_tone_with_five=True, style=Style.INITIALS)
-        orig_finals = g2pw.lazy_pinyin(
-            word, neutral_tone_with_five=True, style=Style.FINALS_TONE3
-        )
 
     for c, v in zip(orig_initials, orig_finals):
         initials.append(c)
@@ -108,6 +85,54 @@ def _get_initials_finals(word):
     return initials, finals
 
 
+must_erhua = {
+    "小院儿", "胡同儿", "范儿", "老汉儿", "撒欢儿", "寻老礼儿", "妥妥儿", "媳妇儿"
+}
+not_erhua = {
+    "虐儿", "为儿", "护儿", "瞒儿", "救儿", "替儿", "有儿", "一儿", "我儿", "俺儿", "妻儿",
+    "拐儿", "聋儿", "乞儿", "患儿", "幼儿", "孤儿", "婴儿", "婴幼儿", "连体儿", "脑瘫儿",
+    "流浪儿", "体弱儿", "混血儿", "蜜雪儿", "舫儿", "祖儿", "美儿", "应采儿", "可儿", "侄儿",
+    "孙儿", "侄孙儿", "女儿", "男儿", "红孩儿", "花儿", "虫儿", "马儿", "鸟儿", "猪儿", "猫儿",
+    "狗儿", "少儿"
+}
+def _merge_erhua(initials: list[str],
+                finals: list[str],
+                word: str,
+                pos: str) -> list[list[str]]:
+    """
+    Do erhub.
+    """
+    # fix er1
+    for i, phn in enumerate(finals):
+        if i == len(finals) - 1 and word[i] == "儿" and phn == 'er1':
+            finals[i] = 'er2'
+
+    # 发音
+    if word not in must_erhua and (word in not_erhua or
+                                        pos in {"a", "j", "nr"}):
+        return initials, finals
+
+    # "……" 等情况直接返回
+    if len(finals) != len(word):
+        return initials, finals
+
+    assert len(finals) == len(word)
+
+    # 与前一个字发同音
+    new_initials = []
+    new_finals = []
+    for i, phn in enumerate(finals):
+        if i == len(finals) - 1 and word[i] == "儿" and phn in {
+                "er2", "er5"
+        } and word[-2:] not in not_erhua and new_finals:
+            phn = "er" + new_finals[-1][-1]
+
+        new_initials.append(initials[i])
+        new_finals.append(phn)
+
+    return new_initials, new_finals
+
+
 def _g2p(segments):
     phones_list = []
     word2ph = []
@@ -116,20 +141,62 @@ def _g2p(segments):
         # Replace all English words in the sentence
         seg = re.sub("[a-zA-Z]+", "", seg)
         seg_cut = psg.lcut(seg)
+        seg_cut = tone_modifier.pre_merge_for_modify(seg_cut)
         initials = []
         finals = []
-        seg_cut = tone_modifier.pre_merge_for_modify(seg_cut)
+
-        for word, pos in seg_cut:
+        if not is_g2pw:
-            if pos == "eng":
+            for word, pos in seg_cut:
-                continue
+                if pos == "eng":
-            sub_initials, sub_finals = _get_initials_finals(word)
+                    continue
-            sub_finals = tone_modifier.modified_tone(word, pos, sub_finals)
+                sub_initials, sub_finals = _get_initials_finals(word)
-            initials.append(sub_initials)
+                sub_finals = tone_modifier.modified_tone(word, pos, sub_finals)
-            finals.append(sub_finals)
+                # 儿化
-            # assert len(sub_initials) == len(sub_finals) == len(word)
+                sub_initials, sub_finals = _merge_erhua(sub_initials, sub_finals, word, pos)
-        initials = sum(initials, [])
+                initials.append(sub_initials)
-        finals = sum(finals, [])
+                finals.append(sub_finals)
-        #
+                # assert len(sub_initials) == len(sub_finals) == len(word)
+            initials = sum(initials, [])
+            finals = sum(finals, [])
+            print("pypinyin结果",initials,finals)
+        else:
+            # g2pw采用整句推理
+            pinyins = g2pw.lazy_pinyin(seg, neutral_tone_with_five=True, style=Style.TONE3)
+
+            pre_word_length = 0
+            for word, pos in seg_cut:
+                sub_initials = []
+                sub_finals = []
+                now_word_length = pre_word_length + len(word)
+
+                if pos == 'eng':
+                    pre_word_length = now_word_length
+                    continue
+
+                word_pinyins = pinyins[pre_word_length:now_word_length]
+
+                # 多音字消歧
+                word_pinyins = correct_pronunciation(word,word_pinyins)
+
+                for pinyin in word_pinyins:
+                    if pinyin[0].isalpha():
+                        sub_initials.append(to_initials(pinyin))
+                        sub_finals.append(to_finals_tone3(pinyin,neutral_tone_with_five=True))
+                    else:
+                        sub_initials.append(pinyin)
+                        sub_finals.append(pinyin)
+
+                pre_word_length = now_word_length
+                sub_finals = tone_modifier.modified_tone(word, pos, sub_finals)
+                # 儿化
+                sub_initials, sub_finals = _merge_erhua(sub_initials, sub_finals, word, pos)
+                initials.append(sub_initials)
+                finals.append(sub_finals)
+
+            initials = sum(initials, [])
+            finals = sum(finals, [])
+            print("g2pw结果",initials,finals)
+
         for c, v in zip(initials, finals):
             raw_pinyin = c + v
             # NOTE: post process for pypinyin outputs
@@ -184,6 +251,24 @@ def _g2p(segments):
     return phones_list, word2ph
 
 
+def replace_punctuation_with_en(text):
+    text = text.replace("嗯", "恩").replace("呣", "母")
+    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"[^\u4e00-\u9fa5A-Za-z" + "".join(punctuation) + r"]+", "", replaced_text
+    )
+
+    return replaced_text
+
+def replace_consecutive_punctuation(text):
+    punctuations = ''.join(re.escape(p) for p in punctuation)
+    pattern = f'([{punctuations}])([{punctuations}])+'
+    result = re.sub(pattern, r'\1', text)
+    return result
+
 def text_normalize(text):
     # https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/zh_normalization
     tx = TextNormalizer()
@@ -196,7 +281,6 @@ def text_normalize(text):
     dest_text = replace_consecutive_punctuation(dest_text)
     return dest_text
 
-
 # 不排除英文的文本格式化
 def mix_text_normalize(text):
     # https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/zh_normalization