zhzluke96 commited on
Commit
d8e7d56
1 Parent(s): ec6a7d0
modules/ChatTTS/example.ipynb DELETED
The diff for this file is too large to render. See raw diff
 
modules/api/impl/base_api.py DELETED
@@ -1,48 +0,0 @@
1
- from pydantic import BaseModel
2
-
3
-
4
- from modules.speaker import speaker_mgr
5
-
6
-
7
- from modules.data import styles_mgr
8
-
9
-
10
- from modules.api import utils as api_utils
11
- from modules.api.Api import APIManager
12
-
13
-
14
- async def list_styles():
15
- return {"message": "ok", "data": styles_mgr.list_items()}
16
-
17
-
18
- async def list_speakers():
19
- return {
20
- "message": "ok",
21
- "data": [spk.to_json() for spk in speaker_mgr.list_speakers()],
22
- }
23
-
24
-
25
- class CreateSpeaker(BaseModel):
26
- seed: int
27
- name: str = ""
28
-
29
-
30
- async def create_speaker(request: CreateSpeaker):
31
- speaker = speaker_mgr.create_speaker(request.seed, request.name)
32
- return {"message": "ok", "data": speaker.to_json()}
33
-
34
-
35
- async def refresh_speakers():
36
- speaker_mgr.refresh_speakers()
37
- return {"message": "ok"}
38
-
39
-
40
- def setup(app: APIManager):
41
- app.get("/v1/styles/list", response_model=api_utils.BaseResponse)(list_styles)
42
- app.get("/v1/speakers/list", response_model=api_utils.BaseResponse)(list_speakers)
43
- app.post("/v1/speaker/create", response_model=api_utils.BaseResponse)(
44
- create_speaker
45
- )
46
- app.post("/v1/speaker/refresh", response_model=api_utils.BaseResponse)(
47
- refresh_speakers
48
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
modules/devices.py DELETED
@@ -1,8 +0,0 @@
1
- import torch
2
-
3
-
4
- def torch_gc():
5
- if torch.cuda.is_available():
6
- with torch.cuda.device("cuda"):
7
- torch.cuda.empty_cache()
8
- torch.cuda.ipc_collect()
 
 
 
 
 
 
 
 
 
modules/normalization.py CHANGED
@@ -143,6 +143,8 @@ def replace_unk_tokens(text):
143
  把不在字典里的字符替换为 " , "
144
  """
145
  chat_tts = models.load_chat_tts()
 
 
146
  tokenizer = chat_tts.pretrain_models["tokenizer"]
147
  vocab = tokenizer.get_vocab()
148
  vocab_set = set(vocab.keys())
 
143
  把不在字典里的字符替换为 " , "
144
  """
145
  chat_tts = models.load_chat_tts()
146
+ if "tokenizer" not in chat_tts.pretrain_models:
147
+ return text
148
  tokenizer = chat_tts.pretrain_models["tokenizer"]
149
  vocab = tokenizer.get_vocab()
150
  vocab_set = set(vocab.keys())
modules/utils/normalization.py DELETED
@@ -1,147 +0,0 @@
1
- from modules.utils.zh_normalization.text_normlization import *
2
-
3
- character_map = {
4
- ":": ",",
5
- ";": ",",
6
- "!": "。",
7
- "(": ",",
8
- ")": ",",
9
- "【": ",",
10
- "】": ",",
11
- "『": ",",
12
- "』": ",",
13
- "「": ",",
14
- "」": ",",
15
- "《": ",",
16
- "》": ",",
17
- "-": ",",
18
- "‘": " ",
19
- "“": " ",
20
- "’": " ",
21
- "”": " ",
22
- ":": ",",
23
- ";": ",",
24
- "!": ".",
25
- "(": ",",
26
- ")": ",",
27
- # '[': ',',
28
- # ']': ',',
29
- ">": ",",
30
- "<": ",",
31
- "-": ",",
32
- }
33
-
34
- character_to_word = {
35
- " & ": " and ",
36
- }
37
-
38
-
39
- def apply_character_to_word(text):
40
- for k, v in character_to_word.items():
41
- text = text.replace(k, v)
42
- return text
43
-
44
-
45
- def apply_character_map(text):
46
- translation_table = str.maketrans(character_map)
47
- return text.translate(translation_table)
48
-
49
-
50
- def insert_spaces_between_uppercase(s):
51
- # 使用正则表达式在每个相邻的大写字母之间插入空格
52
- return re.sub(
53
- r"(?<=[A-Z])(?=[A-Z])|(?<=[a-z])(?=[A-Z])|(?<=[\u4e00-\u9fa5])(?=[A-Z])|(?<=[A-Z])(?=[\u4e00-\u9fa5])",
54
- " ",
55
- s,
56
- )
57
-
58
-
59
- def ensure_suffix(a: str, b: str, c: str):
60
- a = a.strip()
61
- if not a.endswith(b):
62
- a += c
63
- return a
64
-
65
-
66
- email_domain_map = {
67
- "outlook.com": "Out look",
68
- "hotmail.com": "Hot mail",
69
- "yahoo.com": "雅虎",
70
- }
71
-
72
-
73
- # 找到所有 email 并将 name 分割为单个字母,@替换为 at ,. 替换为 dot,常见域名替换为单词
74
- #
75
- # 例如:
76
- # [email protected] => z h z l u k e 9 6 at out look dot com
77
- def email_detect(text):
78
- email_pattern = re.compile(r"([a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,})")
79
-
80
- def replace(match):
81
- email = match.group(1)
82
- name, domain = email.split("@")
83
- name = " ".join(name)
84
- if domain in email_domain_map:
85
- domain = email_domain_map[domain]
86
- domain = domain.replace(".", " dot ")
87
- return f"{name} at {domain}"
88
-
89
- return email_pattern.sub(replace, text)
90
-
91
-
92
- def pre_normalize(text):
93
- # NOTE: 效果一般...
94
- # text = email_detect(text)
95
- return text
96
-
97
-
98
- def post_normalize(text):
99
- text = insert_spaces_between_uppercase(text)
100
- text = apply_character_map(text)
101
- text = apply_character_to_word(text)
102
- return text
103
-
104
-
105
- def text_normalize(text, is_end=False):
106
- # https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/zh_normalization
107
- tx = TextNormalizer()
108
-
109
- # 匹配 \[.+?\] 的部分
110
- pattern = re.compile(r"(\[.+?\])|([^[]+)")
111
-
112
- def normalize_part(part):
113
- part = pre_normalize(part)
114
- sentences = tx.normalize(part)
115
- dest_text = ""
116
- for sentence in sentences:
117
- dest_text += post_normalize(sentence)
118
- return dest_text
119
-
120
- def replace(match):
121
- if match.group(1):
122
- return f" {match.group(1)} "
123
- else:
124
- return normalize_part(match.group(2))
125
-
126
- result = pattern.sub(replace, text)
127
-
128
- # NOTE: 加了会有杂音...
129
- # if is_end:
130
- # 加这个是为了防止吞字
131
- # result = ensure_suffix(result, "[uv_break]", "。。。[uv_break]。。。")
132
-
133
- return result
134
-
135
-
136
- if __name__ == "__main__":
137
- print(
138
- text_normalize(
139
- "ChatTTS是专门为对话场景设计的文本转语音模型,例如LLM助手对话任务。它支持英文和中文两种语言。最大的模型使用了10万小时以上的中英文数据进行训练。在HuggingFace中开源的版本为4万小时训练且未SFT的版本."
140
- )
141
- )
142
- print(
143
- text_normalize(
144
- " [oral_9] [laugh_0] [break_0] 电 [speed_0] 影 [speed_0] 中 梁朝伟 [speed_9] 扮演的陈永仁的编号27149"
145
- )
146
- )
147
- print(text_normalize(" 明天有62%的概率降雨"))