0330-125210-improve_ENG_inference

app.py

@@ -790,14 +790,21 @@ with gr.Blocks(theme='Kasien/ali_theme_custom') as app:
         chinese_choice = gr.Radio(chinese_models, label="ZH",scale=2)
         japanese_choice = gr.Radio(japanese_models, label="JA",scale=4)
 
-    plsh='Support【English/中文/日本語】，Input text you like / 輸入文字 /テキストを入力する'
+    plsh='''
+      Support【English/中文/日本語】，Input text here / 在这輸入文字 /ここにテキストを入力する。
+     
+      If you don't know what to input, you can click the dice on the right, and random text will appear.
+      如果你不知道输入什么，可以点击右边的骰子，会出现随机文本。
+      入力するものがわからない場合は、右側のサイコロをクリックすると、ランダムなテキストが表示されます。
+  
+    '''
     limit='Max 70 words. Excess will be ignored./单次最多处理120字左右，多余的会被忽略'
 
     gr.HTML('''
     <b>Input Text/输入文字</b>''')
     with gr.Row():
         with gr.Column(scale=2): 
-            model_name = gr.Textbox(label="Seleted Model/已选模型", value=default_model_name, scale=1) 
+            model_name = gr.Textbox(label="Seleted Model/已选模型", value=default_model_name, interactive=False,scale=1,) 
             text_language = gr.Textbox(
             label="Language for input text/生成语言",
             info='Automatic detection of input language type.',scale=1,interactive=False
@@ -851,8 +858,12 @@ with gr.Blocks(theme='Kasien/ali_theme_custom') as app:
     If it generated silence, please try again./如果生成了空白声音，请重试
     <br><br><br><br>
     <h1 style="font-size: 25px;">Clone custom Voice/克隆自定义声音</h1>
-    <p style="margin-bottom: 10px; font-size: 100%">Need 3~10s audio.This involves voice-to-text conversion followed by text-to-voice conversion, so it takes longer time<br>
-    需要3~10秒语音，这个会涉及语音转文字，之后再转语音，所以耗时比较久
+    <p style="margin-bottom: 10px; font-size: 100%">
+    需要3~10秒语音，克隆后的声音和原音相似度80%以上<br>
+    Requires 3-10 seconds of voice input. The cloned voice will have a similarity of 80% or above compared to the original.<br>
+    3~10秒の音声入力が必要です。クローンされた音声は、オリジナルと80%以上の類似性があります。
+
+    
     </p>''')
     
     with gr.Row():

gweight.txt

@@ -1 +0,0 @@
-/content/Multi-voice-TTS-GPT-SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt

requirements.txt

@@ -27,4 +27,5 @@ torch
 pyicu 
 morfessor 
 pycld2
-polyglot
+polyglot
+wordsegment

sweight.txt

@@ -1 +0,0 @@
-/content/Multi-voice-TTS-GPT-SoVITS/pretrained_models/s2G488k.pth

text/chinese.py

@@ -30,10 +30,12 @@ rep_map = {
     "\n": ".",
     "·": ",",
     "、": ",",
-    # "...": "…",
+    "...": "…",
     "$": ".",
     "/": ",",
     "—": "-",
+    "~": "…",
+    "～":"…",
 }
 
 tone_modifier = ToneSandhi()
@@ -169,4 +171,4 @@ if __name__ == "__main__":
 
 # # 示例用法
 # text = "这是一个示例文本：,你好！这是一个测试..."
-# print(g2p_paddle(text))  # 输出: 这是一个示例文本你好这是一个测试
+# print(g2p_paddle(text))  # 输出: 这是一个示例文本你好这是一个测试

text/engdict-hot.rep

@@ -1 +1,2 @@
-CHATGPT CH AE1 T JH IY1 P IY1 T IY1
+CHATGPT CH AE1 T JH IY1 P IY1 T IY1
+JSON JH EY1 S AH0 N

text/engdict_cache.pickle

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:36ea4c34f2ca534437a39db0fc4092efa0595093d3587bafc33fc50468ef53f9
-size 6530139
+oid sha256:9bff9393f4b192d873a11335efc8f124771087b6dc847d34fd240c2846889d2b
+size 5965909

text/english.py

@@ -1,18 +1,26 @@
 import pickle
 import os
 import re
+import wordsegment
 from g2p_en import G2p
 
 from string import punctuation
 
 from text import symbols
 
+import unicodedata
+from builtins import str as unicode
+from g2p_en.expand import normalize_numbers
+from nltk.tokenize import TweetTokenizer
+word_tokenize = TweetTokenizer().tokenize
+from nltk import pos_tag
+
 current_file_path = os.path.dirname(__file__)
 CMU_DICT_PATH = os.path.join(current_file_path, "cmudict.rep")
 CMU_DICT_FAST_PATH = os.path.join(current_file_path, "cmudict-fast.rep")
 CMU_DICT_HOT_PATH = os.path.join(current_file_path, "engdict-hot.rep")
 CACHE_PATH = os.path.join(current_file_path, "engdict_cache.pickle")
-_g2p = G2p()
+NAMECACHE_PATH = os.path.join(current_file_path, "namedict_cache.pickle")
 
 arpa = {
     "AH0",
@@ -90,7 +98,7 @@ arpa = {
 
 
 def replace_phs(phs):
-    rep_map = {";": ",", ":": ",", "'": "-", '"': "-"}
+    rep_map = {"'": "-"}
     phs_new = []
     for ph in phs:
         if ph in symbols:
@@ -112,7 +120,7 @@ def read_dict():
             if line_index >= start_line:
                 line = line.strip()
                 word_split = line.split("  ")
-                word = word_split[0]
+                word = word_split[0].lower()
 
                 syllable_split = word_split[1].split(" - ")
                 g2p_dict[word] = []
@@ -132,16 +140,11 @@ def read_dict_new():
         line = f.readline()
         line_index = 1
         while line:
-            if line_index >= 49:
+            if line_index >= 57:
                 line = line.strip()
                 word_split = line.split("  ")
-                word = word_split[0]
-
-                syllable_split = word_split[1].split(" - ")
-                g2p_dict[word] = []
-                for syllable in syllable_split:
-                    phone_split = syllable.split(" ")
-                    g2p_dict[word].append(phone_split)
+                word = word_split[0].lower()
+                g2p_dict[word] = [word_split[1].split(" ")]
 
             line_index = line_index + 1
             line = f.readline()
@@ -153,14 +156,16 @@ def read_dict_new():
             if line_index >= 0:
                 line = line.strip()
                 word_split = line.split(" ")
-                word = word_split[0]
+                word = word_split[0].lower()
                 if word not in g2p_dict:
-                    g2p_dict[word] = []
-                    g2p_dict[word].append(word_split[1:])
+                    g2p_dict[word] = [word_split[1:]]
 
             line_index = line_index + 1
             line = f.readline()
 
+    return g2p_dict
+
+def hot_reload_hot(g2p_dict):
     with open(CMU_DICT_HOT_PATH) as f:
         line = f.readline()
         line_index = 1
@@ -168,14 +173,13 @@ def read_dict_new():
             if line_index >= 0:
                 line = line.strip()
                 word_split = line.split(" ")
-                word = word_split[0]
-                #if word not in g2p_dict:
-                g2p_dict[word] = []
-                g2p_dict[word].append(word_split[1:])
+                word = word_split[0].lower()
+                # 自定义发音词直接覆盖字典
+                g2p_dict[word] = [word_split[1:]]
 
             line_index = line_index + 1
             line = f.readline()
-    
+
     return g2p_dict
 
 
@@ -192,43 +196,167 @@ def get_dict():
         g2p_dict = read_dict_new()
         cache_dict(g2p_dict, CACHE_PATH)
 
+    g2p_dict = hot_reload_hot(g2p_dict)
+
     return g2p_dict
 
 
-eng_dict = get_dict()
+def get_namedict():
+    if os.path.exists(NAMECACHE_PATH):
+        with open(NAMECACHE_PATH, "rb") as pickle_file:
+            name_dict = pickle.load(pickle_file)
+    else:
+        name_dict = {}
+
+    return name_dict
 
 
 def text_normalize(text):
     # todo: eng text normalize
-    return text.replace(";", ",")
+    # 适配中文及 g2p_en 标点
+    rep_map = {
+        "[;:：，；]": ",",
+        '["’]': "'",
+        "。": ".",
+        "！": "!",
+        "？": "?",
+    }
+    for p, r in rep_map.items():
+        text = re.sub(p, r, text)
+
+    # 来自 g2p_en 文本格式化处理
+    # 增加大写兼容
+    text = unicode(text)
+    text = normalize_numbers(text)
+    text = ''.join(char for char in unicodedata.normalize('NFD', text)
+                    if unicodedata.category(char) != 'Mn')  # Strip accents
+    text = re.sub("[^ A-Za-z'.,?!\-]", "", text)
+    text = re.sub(r"(?i)i\.e\.", "that is", text)
+    text = re.sub(r"(?i)e\.g\.", "for example", text)
+
+    return text
+
+
+class en_G2p(G2p):
+    def __init__(self):
+        super().__init__()
+        # 分词初始化
+        wordsegment.load()
+
+        # 扩展过时字典, 添加姓名字典
+        self.cmu = get_dict()
+        self.namedict = get_namedict()
+
+        # 剔除读音错误的几个缩写
+        for word in ["AE", "AI", "AR", "IOS", "HUD", "OS"]:
+            del self.cmu[word.lower()]
+
+        # 修正多音字
+        self.homograph2features["read"] = (['R', 'IY1', 'D'], ['R', 'EH1', 'D'], 'VBP')
+        self.homograph2features["complex"] = (['K', 'AH0', 'M', 'P', 'L', 'EH1', 'K', 'S'], ['K', 'AA1', 'M', 'P', 'L', 'EH0', 'K', 'S'], 'JJ')
+
+
+    def __call__(self, text):
+        # tokenization
+        words = word_tokenize(text)
+        tokens = pos_tag(words)  # tuples of (word, tag)
+
+        # steps
+        prons = []
+        for o_word, pos in tokens:
+            # 还原 g2p_en 小写操作逻辑
+            word = o_word.lower()
+
+            if re.search("[a-z]", word) is None:
+                pron = [word]
+            # 先把单字母推出去
+            elif len(word) == 1:
+                # 单读 A 发音修正, 这里需要原格式 o_word 判断大写
+                if o_word == "A":
+                    pron = ['EY1']
+                else:
+                    pron = self.cmu[word][0]
+            # g2p_en 原版多音字处理
+            elif word in self.homograph2features:  # Check homograph
+                pron1, pron2, pos1 = self.homograph2features[word]
+                if pos.startswith(pos1):
+                    pron = pron1
+                # pos1比pos长仅出现在read
+                elif len(pos) < len(pos1) and pos == pos1[:len(pos)]:
+                    pron = pron1
+                else:
+                    pron = pron2
+            else:
+                # 递归查找预测
+                pron = self.qryword(o_word)
 
+            prons.extend(pron)
+            prons.extend([" "])
 
-def g2p(text):
-    phones = []
-    words = re.split(r"([,;.\-\?\!\s+])", text)
-    for w in words:
-        if w.upper() in eng_dict:
-            phns = eng_dict[w.upper()]
-            for ph in phns:
-                phones += ph
-        else:
-            phone_list = list(filter(lambda p: p != " ", _g2p(w)))
-            for ph in phone_list:
-                if ph in arpa:
-                    phones.append(ph)
+        return prons[:-1]
+
+
+    def qryword(self, o_word):
+        word = o_word.lower()
+
+        # 查字典, 单字母除外
+        if len(word) > 1 and word in self.cmu:  # lookup CMU dict
+            return self.cmu[word][0]
+
+        # 单词仅首字母大写时查找姓名字典
+        if o_word.istitle() and word in self.namedict:
+            return self.namedict[word][0]
+
+        # oov 长度小于等于 3 直接读字母
+        if len(word) <= 3:
+            phones = []
+            for w in word:
+                # 单读 A 发音修正, 此处不存在大写的情况
+                if w == "a":
+                    phones.extend(['EY1'])
                 else:
-                    phones.append(ph)
+                    phones.extend(self.cmu[w][0])
+            return phones
+
+        # 尝试分离所有格
+        if re.match(r"^([a-z]+)('s)$", word):
+            phones = self.qryword(word[:-2])
+            # P T K F TH HH 无声辅音结尾 's 发 ['S']
+            if phones[-1] in ['P', 'T', 'K', 'F', 'TH', 'HH']:
+                phones.extend(['S'])
+            # S Z SH ZH CH JH 擦声结尾 's 发 ['IH1', 'Z'] 或 ['AH0', 'Z']
+            elif phones[-1] in ['S', 'Z', 'SH', 'ZH', 'CH', 'JH']:
+                phones.extend(['AH0', 'Z'])
+            # B D G DH V M N NG L R W Y 有声辅音结尾 's 发 ['Z']
+            # AH0 AH1 AH2 EY0 EY1 EY2 AE0 AE1 AE2 EH0 EH1 EH2 OW0 OW1 OW2 UH0 UH1 UH2 IY0 IY1 IY2 AA0 AA1 AA2 AO0 AO1 AO2
+            # ER ER0 ER1 ER2 UW0 UW1 UW2 AY0 AY1 AY2 AW0 AW1 AW2 OY0 OY1 OY2 IH IH0 IH1 IH2 元音结尾 's 发 ['Z']
+            else:
+                phones.extend(['Z'])
+            return phones
+
+        # 尝试进行分词，应对复合词
+        comps = wordsegment.segment(word.lower())
+
+        # 无法分词的送回去预测
+        if len(comps)==1:
+            return self.predict(word)
+
+        # 可以分词的递归处理
+        return [phone for comp in comps for phone in self.qryword(comp)]
+
+
+_g2p = en_G2p()
+
+
+def g2p(text):
+    # g2p_en 整段推理，剔除不存在的arpa返回
+    phone_list = _g2p(text)
+    phones = [ph if ph != "<unk>" else "UNK" for ph in phone_list if ph not in [" ", "<pad>", "UW", "</s>", "<s>"]]
 
     return replace_phs(phones)
 
 
 if __name__ == "__main__":
-    # print(get_dict())
     print(g2p("hello"))
-    print(g2p("In this; paper, we propose 1 DSPGAN, a GAN-based universal vocoder."))
-    # all_phones = set()
-    # for k, syllables in eng_dict.items():
-    #     for group in syllables:
-    #         for ph in group:
-    #             all_phones.add(ph)
-    # print(all_phones)
+    print(g2p(text_normalize("e.g. I used openai's AI tool to draw a picture.")))
+    print(g2p(text_normalize("In this; paper, we propose 1 DSPGAN, a GAN-based universal vocoder.")))

output_audio.wav → text/namedict_cache.pickle

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6b97115a0643627c7837d43c7a06496cff5cb85f72f2b540d282ded656808700
-size 311084
+oid sha256:559552094c4a6e995213e3fa586330e078ef8cb3a7a95a3109e945111cd2bfc1
+size 760663

text/tone_sandhi.py

@@ -672,6 +672,7 @@ class ToneSandhi:
                 and i + 1 < len(seg)
                 and seg[i - 1][0] == seg[i + 1][0]
                 and seg[i - 1][1] == "v"
+                and seg[i + 1][1] == "v"
             ):
                 new_seg[i - 1][0] = new_seg[i - 1][0] + "一" + new_seg[i - 1][0]
             else: