import logging
logging.getLogger('numba').setLevel(logging.WARNING)
logging.getLogger('matplotlib').setLevel(logging.WARNING)
logging.getLogger('urllib3').setLevel(logging.WARNING)
import json
import re
import numpy as np
import IPython.display as ipd
import torch
import commons
import utils
from models import SynthesizerTrn
from text.symbols import symbols
from text import text_to_sequence
import gradio as gr
import time
import datetime
import os
import pickle
import openai
from scipy.io.wavfile import write
def is_japanese(string):
for ch in string:
if ord(ch) > 0x3040 and ord(ch) < 0x30FF:
return True
return False
def is_english(string):
import re
pattern = re.compile('^[A-Za-z0-9.,:;!?()_*"\' ]+$')
if pattern.fullmatch(string):
return True
else:
return False
def to_html(chat_history):
chat_html = ""
for item in chat_history:
if item['role'] == 'user':
chat_html += f"""
"""
else:
chat_html += f"""
"""
output_html = f"""
{chat_html}
"""
return output_html
def extrac(text):
text = re.sub("<[^>]*>","",text)
result_list = re.split(r'\n', text)
final_list = []
for i in result_list:
if is_english(i):
i = romajitable.to_kana(i).katakana
i = i.replace('\n','').replace(' ','')
#Current length of single sentence: 20
if len(i)>1:
if len(i) > 20:
try:
cur_list = re.split(r'。|！', i)
for i in cur_list:
if len(i)>1:
final_list.append(i+'。')
except:
pass
else:
final_list.append(i)
final_list = [x for x in final_list if x != '']
print(final_list)
return final_list
def to_numpy(tensor: torch.Tensor):
return tensor.detach().cpu().numpy() if tensor.requires_grad \
else tensor.detach().numpy()
def chatgpt(text):
messages = []
try:
with open('log.pickle', 'rb') as f:
messages = pickle.load(f)
messages.append({"role": "user", "content": text},)
chat = openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=messages)
reply = chat.choices[0].message.content
messages.append({"role": "assistant", "content": reply})
print(messages[-1])
if len(messages) == 12:
messages[6:10] = messages[8:]
del messages[-2:]
with open('log.pickle', 'wb') as f:
pickle.dump(messages, f)
return reply,messages
except:
messages.append({"role": "user", "content": text},)
chat = openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=messages)
reply = chat.choices[0].message.content
messages.append({"role": "assistant", "content": reply})
print(messages[-1])
if len(messages) == 12:
messages[6:10] = messages[8:]
del messages[-2:]
with open('log.pickle', 'wb') as f:
pickle.dump(messages, f)
return reply,messages
def get_symbols_from_json(path):
assert os.path.isfile(path)
with open(path, 'r') as f:
data = json.load(f)
return data['symbols']
def sle(language,text):
text = text.replace('\n', ' ').replace('\r', '').replace(" ", "")
if language == "中文":
tts_input1 = "[ZH]" + text + "[ZH]"
return tts_input1
elif language == "自动":
tts_input1 = f"[JA]{text}[JA]" if is_japanese(text) else f"[ZH]{text}[ZH]"
return tts_input1
elif language == "日文":
tts_input1 = "[JA]" + text + "[JA]"
return tts_input1
elif language == "英文":
tts_input1 = "[EN]" + text + "[EN]"
return tts_input1
elif language == "手动":
return text
def get_text(text,hps_ms):
text_norm = text_to_sequence(text,hps_ms.data.text_cleaners)
if hps_ms.data.add_blank:
text_norm = commons.intersperse(text_norm, 0)
text_norm = torch.LongTensor(text_norm)
return text_norm
def create_tts_fn(net_g,hps,speaker_id):
speaker_id = int(speaker_id)
def tts_fn(is_gpt,api_key,is_audio,audiopath,repeat_time,text, language, extract, n_scale= 0.667,n_scale_w = 0.8, l_scale = 1 ):
repeat_ime = int(repeat_time)
if is_gpt:
openai.api_key = api_key
text,messages = chatgpt(text)
htm = to_html(messages)
else:
htm = ''
if not extract:
t1 = time.time()
stn_tst = get_text(sle(language,text),hps)
with torch.no_grad():
x_tst = stn_tst.unsqueeze(0).to(dev)
x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(dev)
sid = torch.LongTensor([speaker_id]).to(dev)
audio = net_g.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=n_scale, noise_scale_w=n_scale_w, length_scale=l_scale)[0][0,0].data.cpu().float().numpy()
t2 = time.time()
spending_time = "推理时间为："+str(t2-t1)+"s"
print(spending_time)
file_path = "subtitles.srt"
try:
write(audiopath + '.wav',22050,audio)
if is_audio:
for i in range(repeat_time):
cmd = 'ffmpeg -y -i ' + audiopath + '.wav' + ' -ar 44100 '+ audiopath.replace('temp','temp'+str(i))
os.system(cmd)
except:
pass
return (hps.data.sampling_rate, audio),file_path,htm
else:
a = ['【','[','(','（']
b = ['】',']',')','）']
for i in a:
text = text.replace(i,'<')
for i in b:
text = text.replace(i,'>')
final_list = extrac(text.replace('“','').replace('”',''))
audio_fin = []
c = 0
t = datetime.timedelta(seconds=0)
for sentence in final_list:
try:
f1 = open("subtitles.srt",'w',encoding='utf-8')
c +=1
stn_tst = get_text(sle(language,sentence),hps)
with torch.no_grad():
x_tst = stn_tst.unsqueeze(0).to(dev)
x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(dev)
sid = torch.LongTensor([speaker_id]).to(dev)
t1 = time.time()
audio = net_g.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=n_scale, noise_scale_w=n_scale_w, length_scale=l_scale)[0][0,0].data.cpu().float().numpy()
t2 = time.time()
spending_time = "第"+str(c)+"句的推理时间为："+str(t2-t1)+"s"
print(spending_time)
time_start = str(t).split(".")[0] + "," + str(t.microseconds)[:3]
last_time = datetime.timedelta(seconds=len(audio)/float(22050))
t+=last_time
time_end = str(t).split(".")[0] + "," + str(t.microseconds)[:3]
print(time_end)
f1.write(str(c-1)+'\n'+time_start+' --> '+time_end+'\n'+sentence+'\n\n')
audio_fin.append(audio)
except:
pass
try:
write(audiopath + '.wav',22050,np.concatenate(audio_fin))
if is_audio:
for i in range(repeat_time):
cmd = 'ffmpeg -y -i ' + audiopath + '.wav' + ' -ar 44100 '+ audiopath.replace('temp','temp'+str(i))
os.system(cmd)
except:
pass
file_path = "subtitles.srt"
return (hps.data.sampling_rate, np.concatenate(audio_fin)),file_path,htm
return tts_fn
if __name__ == '__main__':
hps = utils.get_hparams_from_file('checkpoints/tmp/config.json')
dev = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
models = []
schools = ["Nijigasaki High School","Seisho-Nijigasaki","Seisho Music Academy","Rinmeikan Girls School","Frontier School of Arts","Siegfeld Institute of Music"]
lan = ["中文","日文","自动","手动"]
with open("checkpoints/info.json", "r", encoding="utf-8") as f:
models_info = json.load(f)
for i in models_info:
school = models_info[i]
speakers = school["speakers"]
checkpoint = school["checkpoint"]
phone_dict = {
symbol: i for i, symbol in enumerate(symbols)
}
net_g = SynthesizerTrn(
len(symbols),
hps.data.filter_length // 2 + 1,
hps.train.segment_size // hps.data.hop_length,
n_speakers=hps.data.n_speakers,
**hps.model).to(dev)
_ = net_g.eval()
_ = utils.load_checkpoint(checkpoint, net_g)
content = []
for j in speakers:
sid = int(speakers[j]['sid'])
title = school
example = speakers[j]['speech']
name = speakers[j]["name"]
content.append((sid, name, title, example, create_tts_fn(net_g,hps,sid)))
models.append(content)
with gr.Blocks() as app:
gr.Markdown(
"# Seisho-Nijigasaki vits-models with chatghpt support\n"
"# 少歌&&虹团vits\n"
"## Please do not generate content that could infringe upon the rights or cause harm to individuals or organizations.\n"
"## 请不要生成会对个人以及企划造成侵害的内容\n"
"## 自带chatgpt支持，长句分割支持，srt字幕生成，可修改音频生成路径至live2d语音路径\n"
"## 数据集版权归官方所有，严禁商用及恶意使用"
)
with gr.Tabs():
for i in schools:
with gr.TabItem(i):
for (sid, name, title, example, tts_fn) in models[schools.index(i)]:
with gr.TabItem(name):
with gr.Column():
with gr.Row():
with gr.Row():
gr.Markdown(
''
f'
'
'