语言分割及格式化优化 (#2488)

* better LangSegmenter

* add version num2str

* better version num2str

* sync fast infer

* sync api

* remove duplicate spaces

* remove unnecessary code

---------

Co-authored-by: RVC-Boss <129054828+RVC-Boss@users.noreply.github.com>

GPT_SoVITS/TTS_infer_pack/TextPreprocessor.py

@@ -121,33 +121,31 @@ class TextPreprocessor:
 
     def get_phones_and_bert(self, text: str, language: str, version: str, final: bool = False):
         with self.bert_lock:
-            if language in {"en", "all_zh", "all_ja", "all_ko", "all_yue"}:
-                # language = language.replace("all_","")
-                formattext = text
-                while "  " in formattext:
-                    formattext = formattext.replace("  ", " ")
-                if language == "all_zh":
-                    if re.search(r"[A-Za-z]", formattext):
-                        formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
-                        formattext = chinese.mix_text_normalize(formattext)
-                        return self.get_phones_and_bert(formattext, "zh", version)
-                    else:
-                        phones, word2ph, norm_text = self.clean_text_inf(formattext, language, version)
-                        bert = self.get_bert_feature(norm_text, word2ph).to(self.device)
-                elif language == "all_yue" and re.search(r"[A-Za-z]", formattext):
-                    formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
-                    formattext = chinese.mix_text_normalize(formattext)
-                    return self.get_phones_and_bert(formattext, "yue", version)
-                else:
-                    phones, word2ph, norm_text = self.clean_text_inf(formattext, language, version)
-                    bert = torch.zeros(
-                        (1024, len(phones)),
-                        dtype=torch.float32,
-                    ).to(self.device)
-            elif language in {"zh", "ja", "ko", "yue", "auto", "auto_yue"}:
+            text = re.sub(r' {2,}', ' ', text)
             textlist = []
             langlist = []
-                if language == "auto":
+            if language == "all_zh":
+                for tmp in LangSegmenter.getTexts(text,"zh"):
+                    langlist.append(tmp["lang"])
+                    textlist.append(tmp["text"])
+            elif language == "all_yue":
+                for tmp in LangSegmenter.getTexts(text,"zh"):
+                    if tmp["lang"] == "zh":
+                        tmp["lang"] = "yue"
+                    langlist.append(tmp["lang"])
+                    textlist.append(tmp["text"])
+            elif language == "all_ja":
+                for tmp in LangSegmenter.getTexts(text,"ja"):
+                    langlist.append(tmp["lang"])
+                    textlist.append(tmp["text"])
+            elif language == "all_ko":
+                for tmp in LangSegmenter.getTexts(text,"ko"):
+                    langlist.append(tmp["lang"])
+                    textlist.append(tmp["text"])
+            elif language == "en":
+                langlist.append("en")
+                textlist.append(text)
+            elif language == "auto":
                 for tmp in LangSegmenter.getTexts(text):
                     langlist.append(tmp["lang"])
                     textlist.append(tmp["text"])
@@ -160,9 +158,7 @@ class TextPreprocessor:
             else:
                 for tmp in LangSegmenter.getTexts(text):
                     if langlist:
-                            if (tmp["lang"] == "en" and langlist[-1] == "en") or (
-                                tmp["lang"] != "en" and langlist[-1] != "en"
-                            ):
+                        if (tmp["lang"] == "en" and langlist[-1] == "en") or (tmp["lang"] != "en" and langlist[-1] != "en"):
                             textlist[-1] += tmp["text"]
                             continue
                     if tmp["lang"] == "en":

GPT_SoVITS/inference_webui.py

@@ -586,32 +586,31 @@ from text import chinese
 
 
 def get_phones_and_bert(text, language, version, final=False):
-    if language in {"en", "all_zh", "all_ja", "all_ko", "all_yue"}:
-        formattext = text
-        while "  " in formattext:
-            formattext = formattext.replace("  ", " ")
-        if language == "all_zh":
-            if re.search(r"[A-Za-z]", formattext):
-                formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
-                formattext = chinese.mix_text_normalize(formattext)
-                return get_phones_and_bert(formattext, "zh", version)
-            else:
-                phones, word2ph, norm_text = clean_text_inf(formattext, language, version)
-                bert = get_bert_feature(norm_text, word2ph).to(device)
-        elif language == "all_yue" and re.search(r"[A-Za-z]", formattext):
-            formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
-            formattext = chinese.mix_text_normalize(formattext)
-            return get_phones_and_bert(formattext, "yue", version)
-        else:
-            phones, word2ph, norm_text = clean_text_inf(formattext, language, version)
-            bert = torch.zeros(
-                (1024, len(phones)),
-                dtype=torch.float16 if is_half == True else torch.float32,
-            ).to(device)
-    elif language in {"zh", "ja", "ko", "yue", "auto", "auto_yue"}:
+    text = re.sub(r' {2,}', ' ', text)
     textlist = []
     langlist = []
-        if language == "auto":
+    if language == "all_zh":
+        for tmp in LangSegmenter.getTexts(text,"zh"):
+            langlist.append(tmp["lang"])
+            textlist.append(tmp["text"])
+    elif language == "all_yue":
+        for tmp in LangSegmenter.getTexts(text,"zh"):
+            if tmp["lang"] == "zh":
+                tmp["lang"] = "yue"
+            langlist.append(tmp["lang"])
+            textlist.append(tmp["text"])
+    elif language == "all_ja":
+        for tmp in LangSegmenter.getTexts(text,"ja"):
+            langlist.append(tmp["lang"])
+            textlist.append(tmp["text"])
+    elif language == "all_ko":
+        for tmp in LangSegmenter.getTexts(text,"ko"):
+            langlist.append(tmp["lang"])
+            textlist.append(tmp["text"])
+    elif language == "en":
+        langlist.append("en")
+        textlist.append(text)
+    elif language == "auto":
         for tmp in LangSegmenter.getTexts(text):
             langlist.append(tmp["lang"])
             textlist.append(tmp["text"])

GPT_SoVITS/text/LangSegmenter/langsegmenter.py

@@ -3,25 +3,19 @@ import re
 
 # jieba静音
 import jieba
-
 jieba.setLogLevel(logging.CRITICAL)
 
 # 更改fast_langdetect大模型位置
 from pathlib import Path
 import fast_langdetect
-
-fast_langdetect.infer._default_detector = fast_langdetect.infer.LangDetector(
-    fast_langdetect.infer.LangDetectConfig(
-        cache_dir=Path(__file__).parent.parent.parent / "pretrained_models" / "fast_langdetect"
-    )
-)
+fast_langdetect.infer._default_detector = fast_langdetect.infer.LangDetector(fast_langdetect.infer.LangDetectConfig(cache_dir=Path(__file__).parent.parent.parent / "pretrained_models" / "fast_langdetect"))
 
 
 from split_lang import LangSplitter
 
 
 def full_en(text):
-    pattern = r"^(?=.*[A-Za-z])[A-Za-z0-9\s\u0020-\u007E\u2000-\u206F\u3000-\u303F\uFF00-\uFFEF]+$"
+    pattern = r'^(?=.*[A-Za-z])[A-Za-z0-9\s\u0020-\u007E\u2000-\u206F\u3000-\u303F\uFF00-\uFFEF]+$'
     return bool(re.match(pattern, text))
 
 
@@ -40,7 +34,7 @@ def full_cjk(text):
         (0x2EBF0, 0x2EE5D),      # CJK Extension H
     ]
 
-    pattern = r"[0-9、-〜。！？.!?… /]+$"
+    pattern = r'[0-9、-〜。！？.!?… /]+$'
 
     cjk_text = ""
     for char in text:
@@ -59,28 +53,28 @@ def split_jako(tag_lang, item):
 
     lang_list: list[dict] = []
     tag = 0
-    for match in re.finditer(pattern, item["text"]):
+    for match in re.finditer(pattern, item['text']):
         if match.start() > tag:
-            lang_list.append({"lang": item["lang"], "text": item["text"][tag : match.start()]})
+            lang_list.append({'lang':item['lang'],'text':item['text'][tag:match.start()]})
 
         tag = match.end()
-        lang_list.append({"lang": tag_lang, "text": item["text"][match.start() : match.end()]})
+        lang_list.append({'lang':tag_lang,'text':item['text'][match.start():match.end()]})
 
-    if tag < len(item["text"]):
-        lang_list.append({"lang": item["lang"], "text": item["text"][tag : len(item["text"])]})
+    if tag < len(item['text']):
+        lang_list.append({'lang':item['lang'],'text':item['text'][tag:len(item['text'])]})
 
     return lang_list
 
 
 def merge_lang(lang_list, item):
-    if lang_list and item["lang"] == lang_list[-1]["lang"]:
-        lang_list[-1]["text"] += item["text"]
+    if lang_list and item['lang'] == lang_list[-1]['lang']:
+        lang_list[-1]['text'] += item['text']
     else:
         lang_list.append(item)
     return lang_list
 
 
-class LangSegmenter:
+class LangSegmenter():
     # 默认过滤器, 基于gsv目前四种语言
     DEFAULT_LANG_MAP = {
         "zh": "zh",
@@ -93,25 +87,41 @@ class LangSegmenter:
         "en": "en",
     }
 
-    def getTexts(text):
+    def getTexts(text,default_lang = ""):
         lang_splitter = LangSplitter(lang_map=LangSegmenter.DEFAULT_LANG_MAP)
+        lang_splitter.merge_across_digit = False
         substr = lang_splitter.split_by_lang(text=text)
 
         lang_list: list[dict] = []
 
-        for _, item in enumerate(substr):
-            dict_item = {"lang": item.lang, "text": item.text}
+        have_num = False
 
-            # 处理短英文被识别为其他语言的问题
-            if full_en(dict_item["text"]):
-                dict_item["lang"] = "en"
+        for _, item in enumerate(substr):
+            dict_item = {'lang':item.lang,'text':item.text}
+
+            if dict_item['lang'] == 'digit':
+                if default_lang != "":
+                    dict_item['lang'] = default_lang
+                else:
+                    have_num = True
                 lang_list = merge_lang(lang_list,dict_item)
                 continue
 
+            # 处理短英文被识别为其他语言的问题
+            if full_en(dict_item['text']):  
+                dict_item['lang'] = 'en'
+                lang_list = merge_lang(lang_list,dict_item)
+                continue
+
+            if default_lang != "":
+                dict_item['lang'] = default_lang
+                lang_list = merge_lang(lang_list,dict_item)
+                continue
+            else:
                 # 处理非日语夹日文的问题(不包含CJK)
                 ja_list: list[dict] = []
-            if dict_item["lang"] != "ja":
-                ja_list = split_jako("ja", dict_item)
+                if dict_item['lang'] != 'ja':
+                    ja_list = split_jako('ja',dict_item)
 
                 if not ja_list:
                     ja_list.append(dict_item)
@@ -120,8 +130,8 @@ class LangSegmenter:
                 ko_list: list[dict] = []
                 temp_list: list[dict] = []
                 for _, ko_item in enumerate(ja_list):
-                if ko_item["lang"] != "ko":
-                    ko_list = split_jako("ko", ko_item)
+                    if ko_item["lang"] != 'ko':
+                        ko_list = split_jako('ko',ko_item)
 
                     if ko_list:
                         temp_list.extend(ko_list)
@@ -131,10 +141,10 @@ class LangSegmenter:
                 # 未存在非日韩文夹日韩文
                 if len(temp_list) == 1:
                     # 未知语言检查是否为CJK
-                if dict_item["lang"] == "x":
-                    cjk_text = full_cjk(dict_item["text"])
+                    if dict_item['lang'] == 'x':
+                        cjk_text = full_cjk(dict_item['text'])
                         if cjk_text:
-                        dict_item = {"lang": "zh", "text": cjk_text}
+                            dict_item = {'lang':'zh','text':cjk_text}
                             lang_list = merge_lang(lang_list,dict_item)
                         else:
                             lang_list = merge_lang(lang_list,dict_item)
@@ -146,26 +156,57 @@ class LangSegmenter:
                 # 存在非日韩文夹日韩文
                 for _, temp_item in enumerate(temp_list):
                     # 未知语言检查是否为CJK
-                if temp_item["lang"] == "x":
-                    cjk_text = full_cjk(dict_item["text"])
+                    if temp_item['lang'] == 'x':
+                        cjk_text = full_cjk(temp_item['text'])
                         if cjk_text:
-                        dict_item = {"lang": "zh", "text": cjk_text}
-                        lang_list = merge_lang(lang_list, dict_item)
+                            lang_list = merge_lang(lang_list,{'lang':'zh','text':cjk_text})
                         else:
-                        lang_list = merge_lang(lang_list, dict_item)
+                            lang_list = merge_lang(lang_list,temp_item)
                     else:
                         lang_list = merge_lang(lang_list,temp_item)
 
+        # 有数字
+        if have_num:
+            temp_list = lang_list
+            lang_list = []
+            for i, temp_item in enumerate(temp_list):
+                if temp_item['lang'] == 'digit':
+                    if default_lang:
+                        temp_item['lang'] = default_lang
+                    elif lang_list and i == len(temp_list) - 1:
+                        temp_item['lang'] = lang_list[-1]['lang']
+                    elif not lang_list and i < len(temp_list) - 1:
+                        temp_item['lang'] = temp_list[1]['lang']
+                    elif lang_list and i < len(temp_list) - 1:
+                        if lang_list[-1]['lang'] == temp_list[i + 1]['lang']:
+                            temp_item['lang'] = lang_list[-1]['lang']
+                        elif lang_list[-1]['text'][-1] in [",",".","!","?","，","。","！","？"]:
+                            temp_item['lang'] = temp_list[i + 1]['lang']
+                        elif temp_list[i + 1]['text'][0] in [",",".","!","?","，","。","！","？"]:
+                            temp_item['lang'] = lang_list[-1]['lang']
+                        elif temp_item['text'][-1] in ["。","."]:
+                            temp_item['lang'] = lang_list[-1]['lang']
+                        elif len(lang_list[-1]['text']) >= len(temp_list[i + 1]['text']):
+                            temp_item['lang'] = lang_list[-1]['lang']
+                        else:
+                            temp_item['lang'] = temp_list[i + 1]['lang']
+                    else:
+                        temp_item['lang'] = 'zh'
+
+                lang_list = merge_lang(lang_list,temp_item)
+
+
+        # 筛X
         temp_list = lang_list
         lang_list = []
         for _, temp_item in enumerate(temp_list):
-            if temp_item["lang"] == "x":
+            if temp_item['lang'] == 'x':
                 if lang_list:
-                    temp_item["lang"] = lang_list[-1]["lang"]
+                    temp_item['lang'] = lang_list[-1]['lang']
                 elif len(temp_list) > 1:
-                    temp_item["lang"] = temp_list[1]["lang"]
+                    temp_item['lang'] = temp_list[1]['lang']
                 else:
-                    temp_item["lang"] = "zh"
+                    temp_item['lang'] = 'zh'
 
             lang_list = merge_lang(lang_list,temp_item)
 
@@ -178,3 +219,7 @@ if __name__ == "__main__":
 
     text = "ねえ、知ってる？最近、僕は天文学を勉強してるんだ。君の瞳が星空みたいにキラキラしてるからさ。"
     print(LangSegmenter.getTexts(text))
+
+    text = "当时ThinkPad T60刚刚发布，一同推出的还有一款名为Advanced Dock的扩展坞配件。这款扩展坞通过连接T60底部的插槽，扩展出包括PCIe在内的一大堆接口，并且自带电源，让T60可以安装桌面显卡来提升性能。"
+    print(LangSegmenter.getTexts(text,"zh"))
+    print(LangSegmenter.getTexts(text))

GPT_SoVITS/text/chinese.py

@@ -181,20 +181,6 @@ def text_normalize(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
-    tx = TextNormalizer()
-    sentences = tx.normalize(text)
-    dest_text = ""
-    for sentence in sentences:
-        dest_text += replace_punctuation_with_en(sentence)
-
-    # 避免重复标点引起的参考泄露
-    dest_text = replace_consecutive_punctuation(dest_text)
-    return dest_text
-
-
 if __name__ == "__main__":
     text = "啊——但是《原神》是由,米哈\游自主，研发的一款全.新开放世界.冒险游戏"
     text = "呣呣呣～就是…大人的鼹鼠党吧？"

GPT_SoVITS/text/chinese2.py

@@ -326,20 +326,6 @@ def text_normalize(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
-    tx = TextNormalizer()
-    sentences = tx.normalize(text)
-    dest_text = ""
-    for sentence in sentences:
-        dest_text += replace_punctuation_with_en(sentence)
-
-    # 避免重复标点引起的参考泄露
-    dest_text = replace_consecutive_punctuation(dest_text)
-    return dest_text
-
-
 if __name__ == "__main__":
     text = "啊——但是《原神》是由,米哈\游自主，研发的一款全.新开放世界.冒险游戏"
     text = "呣呣呣～就是…大人的鼹鼠党吧？"

GPT_SoVITS/text/zh_normalization/num.py

@@ -256,6 +256,24 @@ def replace_to_range(match) -> str:
     return result
 
 
+RE_VERSION_NUM = re.compile(r"((\d+)(\.\d+)(\.\d+)?(\.\d+)+)")
+def replace_vrsion_num(match) -> str:
+    """
+    Args:
+        match (re.Match)
+    Returns:
+        str
+    """
+    result = ""
+    for c in match.group(1):
+        if c == ".":
+            result += "点"
+        else:
+            result += num2str(c)
+    return result
+
+
+
 def _get_value(value_string: str, use_zero: bool = True) -> List[str]:
     stripped = value_string.lstrip("0")
     if len(stripped) == 0:
@@ -308,7 +326,11 @@ def num2str(value_string: str) -> str:
 
     result = verbalize_cardinal(integer)
 
+    if decimal.endswith("0"):
+        decimal = decimal.rstrip("0") + "0"
+    else:
         decimal = decimal.rstrip("0")
+
     if decimal:
         # '.22' is verbalized as '零点二二'
         # '3.20' is verbalized as '三点二

GPT_SoVITS/text/zh_normalization/text_normlization.py

@@ -25,6 +25,7 @@ from .chronology import replace_time
 from .constants import F2H_ASCII_LETTERS
 from .constants import F2H_DIGITS
 from .constants import F2H_SPACE
+from .num import RE_VERSION_NUM
 from .num import RE_DECIMAL_NUM
 from .num import RE_DEFAULT_NUM
 from .num import RE_FRAC
@@ -36,6 +37,7 @@ from .num import RE_RANGE
 from .num import RE_TO_RANGE
 from .num import RE_ASMD
 from .num import RE_POWER
+from .num import replace_vrsion_num
 from .num import replace_default_num
 from .num import replace_frac
 from .num import replace_negative_num
@@ -158,6 +160,7 @@ class TextNormalizer:
         sentence = RE_RANGE.sub(replace_range, sentence)
 
         sentence = RE_INTEGER.sub(replace_negative_num, sentence)
+        sentence = RE_VERSION_NUM.sub(replace_vrsion_num, sentence)
         sentence = RE_DECIMAL_NUM.sub(replace_number, sentence)
         sentence = RE_POSITIVE_QUANTIFIERS.sub(replace_positive_quantifier, sentence)
         sentence = RE_DEFAULT_NUM.sub(replace_default_num, sentence)

api.py

@@ -543,32 +543,31 @@ from text import chinese
 
 
 def get_phones_and_bert(text, language, version, final=False):
-    if language in {"en", "all_zh", "all_ja", "all_ko", "all_yue"}:
-        formattext = text
-        while "  " in formattext:
-            formattext = formattext.replace("  ", " ")
-        if language == "all_zh":
-            if re.search(r"[A-Za-z]", formattext):
-                formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
-                formattext = chinese.mix_text_normalize(formattext)
-                return get_phones_and_bert(formattext, "zh", version)
-            else:
-                phones, word2ph, norm_text = clean_text_inf(formattext, language, version)
-                bert = get_bert_feature(norm_text, word2ph).to(device)
-        elif language == "all_yue" and re.search(r"[A-Za-z]", formattext):
-            formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
-            formattext = chinese.mix_text_normalize(formattext)
-            return get_phones_and_bert(formattext, "yue", version)
-        else:
-            phones, word2ph, norm_text = clean_text_inf(formattext, language, version)
-            bert = torch.zeros(
-                (1024, len(phones)),
-                dtype=torch.float16 if is_half == True else torch.float32,
-            ).to(device)
-    elif language in {"zh", "ja", "ko", "yue", "auto", "auto_yue"}:
+    text = re.sub(r' {2,}', ' ', text)
     textlist = []
     langlist = []
-        if language == "auto":
+    if language == "all_zh":
+        for tmp in LangSegmenter.getTexts(text,"zh"):
+            langlist.append(tmp["lang"])
+            textlist.append(tmp["text"])
+    elif language == "all_yue":
+        for tmp in LangSegmenter.getTexts(text,"zh"):
+            if tmp["lang"] == "zh":
+                tmp["lang"] = "yue"
+            langlist.append(tmp["lang"])
+            textlist.append(tmp["text"])
+    elif language == "all_ja":
+        for tmp in LangSegmenter.getTexts(text,"ja"):
+            langlist.append(tmp["lang"])
+            textlist.append(tmp["text"])
+    elif language == "all_ko":
+        for tmp in LangSegmenter.getTexts(text,"ko"):
+            langlist.append(tmp["lang"])
+            textlist.append(tmp["text"])
+    elif language == "en":
+        langlist.append("en")
+        textlist.append(text)
+    elif language == "auto":
         for tmp in LangSegmenter.getTexts(text):
             langlist.append(tmp["lang"])
             textlist.append(tmp["text"])