Update app.py

app.py

@@ -2,7 +2,7 @@ import logging
 logging.getLogger('numba').setLevel(logging.WARNING)
 logging.getLogger('matplotlib').setLevel(logging.WARNING)
 logging.getLogger('urllib3').setLevel(logging.WARNING)
-import romajitable
+import json
 import re
 import numpy as np
 import IPython.display as ipd
@@ -16,251 +16,126 @@ import gradio as gr
 import time
 import datetime
 import os
+import pickle
+import openai
+from scipy.io.wavfile import write
 import librosa
 from mel_processing import spectrogram_torch
-class VitsGradio:
-    def __init__(self):
-        self.dev = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-        self.lan = ["中文","日文","自动","手动"]
-        self.idols = ["c1","c2","高咲侑","歩夢","かすみ","しずく","果林","愛","彼方","せつ菜","璃奈","栞子","エマ","ランジュ","ミア","華恋","まひる","なな","クロディーヌ","ひかり",'純那',"香子","真矢","双葉","ミチル","メイファン","やちよ","晶","いちえ","ゆゆ子","塁","珠緒","あるる","ララフィン","美空","静羽","あるる"]
-        self.modelPaths = []
-        for root,dirs,files in os.walk("checkpoints"):
-            for dir in dirs:
-                self.modelPaths.append(dir)
-        with gr.Blocks() as self.Vits:
-            gr.Markdown(
-            "## <center> Lovelive虹团中日双语VITS\n"
-            "### <center> 请不要生成会对个人以及企划造成侵害的内容\n"
-            "<div align='center'>目前有标贝普通话版，去标贝版，少歌模型还是大饼状态</div>"
-            '<div align="center"><a>参数说明:由于爱抖露们过于有感情，合成日语时建议将噪声比例调节至0.2-0.3区间，噪声偏差对应着每个字之间的间隔，对普通话影响较大，duration代表整体语速</div>'
-            '<div align="center"><a>合成前请先选择模型，否则第一次合成不一定成功。长段落/小说合成建议colab或本地运行</div>')
-            with gr.Tab("TTS合成"):
-                with gr.Row():
-                    with gr.Column():
-                        with gr.Row():
-                            with gr.Column():
-                                input1 = gr.TextArea(label="Text", value="为什么你会那么熟练啊？你和雪菜亲过多少次了")
-                                input2 = gr.Dropdown(label="Language", choices=self.lan, value="自动", interactive=True)
-                                input3 = gr.Dropdown(label="Speaker", choices=self.idols, value="歩夢", interactive=True)
-                                btnVC = gr.Button("Submit")
-                            with gr.Column():
-                                input4 = gr.Slider(minimum=0, maximum=1.0, label="更改噪声比例(noise scale)，以控制情感", value=0.267)
-                                input5 = gr.Slider(minimum=0, maximum=1.0, label="更改噪声偏差(noise scale w)，以控制音素长短", value=0.7)
-                                input6 = gr.Slider(minimum=0.1, maximum=10, label="duration", value=1)
-                                output1 = gr.Audio(label="采样率22050")
-                btnVC.click(self.infer, inputs=[input1, input2, input3, input4, input5, input6], outputs=[output1])
-            with gr.Tab("选择模型"):
-                with gr.Column():
-                    modelstrs = gr.Dropdown(label = "模型", choices = self.modelPaths, value = self.modelPaths[0], type = "value")
-                    btnMod = gr.Button("载入模型")
-                    statusa = gr.TextArea()
-                    btnMod.click(self.loadCk, inputs=[modelstrs], outputs = [statusa])
-            with gr.Tab("Voice Conversion"):
-                gr.Markdown("""
-                                录制或上传声音，并选择要转换的音色。
-                """)
-                with gr.Column():
-                    record_audio = gr.Audio(label="record your voice", source="microphone")
-                    upload_audio = gr.Audio(label="or upload audio here", source="upload")
-                    source_speaker = gr.Dropdown(choices=self.idols, value="歩夢", label="source speaker")
-                    target_speaker = gr.Dropdown(choices=self.idols, value="歩夢", label="target speaker")
-                with gr.Column():
-                    message_box = gr.Textbox(label="Message")
-                    converted_audio = gr.Audio(label='converted audio')
-                btn = gr.Button("Convert!")
-                btn.click(self.vc_fn, inputs=[source_speaker, target_speaker, record_audio, upload_audio],
-                        outputs=[message_box, converted_audio])
-            with gr.Tab("小说合成(带字幕)"):
-                with gr.Row():
-                    with gr.Column():
-                        with gr.Row():
-                            with gr.Column():
-                                input1 = gr.TextArea(label="建议colab或本地克隆后运行本仓库", value="为什么你会那么熟练啊？你和雪菜亲过多少次了")
-                                input2 = gr.Dropdown(label="Language", choices=self.lan, value="自动", interactive=True)
-                                input3 = gr.Dropdown(label="Speaker", choices=self.idols, value="歩夢", interactive=True)
-                                btnVC = gr.Button("Submit")
-                            with gr.Column():
-                                input4 = gr.Slider(minimum=0, maximum=1.0, label="更改噪声比例(noise scale)，以控制情感", value=0.267)
-                                input5 = gr.Slider(minimum=0, maximum=1.0, label="更改噪声偏差(noise scale w)，以控制音素长短", value=0.7)
-                                input6 = gr.Slider(minimum=0.1, maximum=10, label="Duration", value=1)
-                                output1 = gr.Audio(label="采样率22050")
-                                subtitle = gr.outputs.File(label="字幕文件：subtitles.srt")
-                btnVC.click(self.infer2, inputs=[input1, input2, input3, input4, input5, input6], outputs=[output1,subtitle])
-    
-    def loadCk(self,path):
-        self.hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
-        self.net_g = SynthesizerTrn(
-            len(symbols),
-            self.hps.data.filter_length // 2 + 1,
-            self.hps.train.segment_size // self.hps.data.hop_length,
-            n_speakers=self.hps.data.n_speakers,
-            **self.hps.model).to(self.dev)
-        _ = self.net_g.eval()
-        _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", self.net_g)
-        return "success"
-
-    def get_text(self,text):
-        text_norm = text_to_sequence(text,self.hps.data.text_cleaners)
-        if self.hps.data.add_blank:
-            text_norm = commons.intersperse(text_norm, 0)
-        text_norm = torch.LongTensor(text_norm)
-        return text_norm
-    
-    def is_japanese(self,string):
+def is_japanese(string):
         for ch in string:
             if ord(ch) > 0x3040 and ord(ch) < 0x30FF:
                 return True
         return False
-    
-    def is_english(self,string):
+
+def is_english(string):
         import re
         pattern = re.compile('^[A-Za-z0-9.,:;!?()_*"\' ]+$')
         if pattern.fullmatch(string):
             return True
         else:
             return False
-    
-    def selection(self,speaker):
-        if speaker == "高咲侑":
-            spk = 0
-            return spk
-
-        elif speaker == "歩夢":
-            spk = 1
-            return spk
-
-        elif speaker == "かすみ":
-            spk = 2
-            return spk
 
-        elif speaker == "しずく":
-            spk = 3
-            return spk
-
-        elif speaker == "果林":
-            spk = 4
-            return spk
-    
-        elif speaker == "愛":
-            spk = 5
-            return spk
+def to_html(chat_history):
+    chat_html = ""
+    for item in chat_history:
+        if item['role'] == 'user':
+            chat_html += f"""
+                <div style="margin-bottom: 20px;">
+                    <div style="text-align: right; margin-right: 20px;">
+                        <span style="background-color: #4CAF50; color: black; padding: 10px; border-radius: 10px; display: inline-block; max-width: 80%; word-wrap: break-word;">
+                            {item['content']}
+                        </span>
+                    </div>
+                </div>
+            """
+        else:
+            chat_html += f"""
+                <div style="margin-bottom: 20px;">
+                    <div style="text-align: left; margin-left: 20px;">
+                        <span style="background-color: white; color: black; padding: 10px; border-radius: 10px; display: inline-block; max-width: 80%; word-wrap: break-word;">
+                            {item['content']}
+                        </span>
+                    </div>
+                </div>
+            """
+    output_html = f"""
+        <div style="height: 400px; overflow-y: scroll; padding: 10px;">
+            {chat_html}
+        </div>
+    """
+    return output_html
 
-        elif speaker == "彼方":
-            spk = 6
-            return spk
+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
 
-        elif speaker == "せつ菜":
-            spk = 7
-            return spk
-        elif speaker == "エマ":
-            spk = 8
-            return spk
-        elif speaker == "璃奈":
-            spk = 9
-            return spk
-        elif speaker == "栞子":
-            spk = 10
-            return spk
-        elif speaker == "ランジュ":
-            spk = 11
-            return spk
-        elif speaker == "ミア":
-            spk = 12
-            return spk
-        
-        elif speaker == "派蒙":
-            spk = 16
-            return spk
-        
-        elif speaker == "c1":
-            spk = 18
-            return spk
+def to_numpy(tensor: torch.Tensor):
+    return tensor.detach().cpu().numpy() if tensor.requires_grad \
+        else tensor.detach().numpy()
 
-        elif speaker == "c2":
-            spk = 19
-            return spk
+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:
+                messages2 = []
+                pickle.dump(messages2, 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
 
-        elif speaker == "華恋":
-            spk = 21
-            return spk
+def get_symbols_from_json(path):
+    assert os.path.isfile(path)
+    with open(path, 'r') as f:
+        data = json.load(f)
+    return data['symbols']
 
-        elif speaker == "まひる":
-            spk = 22
-            return spk
-    
-        elif speaker == "なな":
-            spk = 23
-            return spk
-    
-        elif speaker == "クロディーヌ":
-            spk = 24
-            return spk
-    
-        elif speaker == "ひかり":
-            spk = 25
-            return spk
-    
-        elif speaker == "純那":
-            spk = 26
-            return spk
-    
-        elif speaker == "香子":
-            spk = 27
-            return spk
-    
-        elif speaker == "真矢":
-            spk = 28
-            return spk
-        elif speaker == "双葉":
-            spk = 29
-            return spk
-        elif speaker == "ミチル":
-            spk = 30
-            return spk
-        elif speaker == "メイファン":
-            spk = 31
-            return spk
-        elif speaker == "やちよ":
-            spk = 32
-            return spk
-        elif speaker == "晶":
-            spk = 33
-            return spk
-        elif speaker == "いちえ":
-            spk = 34
-            return spk
-        elif speaker == "ゆゆ子":
-            spk = 35
-            return spk
-        elif speaker == "塁":
-            spk = 36
-            return spk
-        elif speaker == "珠緒":
-            spk = 37
-            return spk
-        elif speaker == "あるる":
-            spk = 38
-            return spk
-        elif speaker == "ララフィン":
-            spk = 39
-            return spk
-        elif speaker == "美空":
-            spk = 40
-            return spk
-        elif speaker == "静羽":
-            spk = 41
-            return spk
-        else:
-            return 0
-            
-    
-    def sle(self,language,text):
-        text = text.replace('\n','。').replace(' ',',')
+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 self.is_japanese(text) else f"[ZH]{text}[ZH]"
+            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]"
@@ -270,119 +145,361 @@ class VitsGradio:
             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 vc_fn(original_speaker, target_speaker, record_audio, upload_audio):
+    input_audio = record_audio if record_audio is not None else upload_audio
+    if input_audio is None:
+        return "You need to record or upload an audio", None
+    sampling_rate, audio = input_audio
+    original_speaker_id = selection(original_speaker)
+    target_speaker_id = selection(target_speaker)
+
+    audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
+    if len(audio.shape) > 1:
+        audio = librosa.to_mono(audio.transpose(1, 0))
+    if sampling_rate != hps.data.sampling_rate:
+        audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=hps.data.sampling_rate)
+    with torch.no_grad():
+        y = torch.FloatTensor(audio)
+        y = y / max(-y.min(), y.max()) / 0.99
+        y = y.to(dev)
+        y = y.unsqueeze(0)
+        spec = spectrogram_torch(y, hps.data.filter_length,
+                                    hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length,
+                                    center=False).to(dev)
+        spec_lengths = torch.LongTensor([spec.size(-1)]).to(dev)
+        sid_src = torch.LongTensor([original_speaker_id]).to(dev)
+        sid_tgt = torch.LongTensor([target_speaker_id]).to(dev)
+        audio = net_g.voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt)[0][
+            0, 0].data.cpu().float().numpy()
+    del y, spec, spec_lengths, sid_src, sid_tgt
+    return "Success", (hps.data.sampling_rate, audio)
+
+def selection(speaker):
+    if speaker == "高咲侑":
+        spk = 0
+        return spk
+
+    elif speaker == "歩夢":
+        spk = 1
+        return spk
+
+    elif speaker == "かすみ":
+        spk = 2
+        return spk
+
+    elif speaker == "しずく":
+        spk = 3
+        return spk
+
+    elif speaker == "果林":
+        spk = 4
+        return spk
+
+    elif speaker == "愛":
+        spk = 5
+        return spk
+
+    elif speaker == "彼方":
+        spk = 6
+        return spk
+
+    elif speaker == "せつ菜":
+        spk = 7
+        return spk
+    
+    elif speaker == "エマ":
+        spk = 8
+        return spk
+    
+    elif speaker == "璃奈":
+        spk = 9
+        return spk
+    
+    elif speaker == "栞子":
+        spk = 10
+        return spk
+    
+    elif speaker == "ランジュ":
+        spk = 11
+        return spk
     
-    def extrac(self,text):
-        text = re.sub("<[^>]*>","",text)
-        result_list = re.split(r'\n', text)
-        final_list = []
-        for i in result_list:
-            if self.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
+    elif speaker == "ミア":
+        spk = 12
+        return spk
     
-    def vc_fn(self,original_speaker, target_speaker, record_audio, upload_audio):
-        input_audio = record_audio if record_audio is not None else upload_audio
-        if input_audio is None:
-            return "You need to record or upload an audio", None
-        sampling_rate, audio = input_audio
-        original_speaker_id = self.selection(original_speaker)
-        target_speaker_id = self.selection(target_speaker)
+    elif speaker == "派蒙":
+        spk = 16
+        return spk
+    
+    elif speaker == "c1":
+        spk = 18
+        return spk
+
+    elif speaker == "c2":
+        spk = 19
+        return spk
+
+    elif speaker == "華恋":
+        spk = 21
+        return spk
+
+    elif speaker == "まひる":
+        spk = 22
+        return spk
+
+    elif speaker == "なな":
+        spk = 23
+        return spk
+
+    elif speaker == "クロディーヌ":
+        spk = 24
+        return spk
 
-        audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
-        if len(audio.shape) > 1:
-            audio = librosa.to_mono(audio.transpose(1, 0))
-        if sampling_rate != self.hps.data.sampling_rate:
-            audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=self.hps.data.sampling_rate)
-        with torch.no_grad():
-            y = torch.FloatTensor(audio)
-            y = y / max(-y.min(), y.max()) / 0.99
-            y = y.to(self.dev)
-            y = y.unsqueeze(0)
-            spec = spectrogram_torch(y, self.hps.data.filter_length,
-                                     self.hps.data.sampling_rate, self.hps.data.hop_length, self.hps.data.win_length,
-                                     center=False).to(self.dev)
-            spec_lengths = torch.LongTensor([spec.size(-1)]).to(self.dev)
-            sid_src = torch.LongTensor([original_speaker_id]).to(self.dev)
-            sid_tgt = torch.LongTensor([target_speaker_id]).to(self.dev)
-            audio = self.net_g.voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt)[0][
-                0, 0].data.cpu().float().numpy()
-        del y, spec, spec_lengths, sid_src, sid_tgt
-        return "Success", (self.hps.data.sampling_rate, audio)
+    elif speaker == "ひかり":
+        spk = 25
+        return spk
+
+    elif speaker == "純那":
+        spk = 26
+        return spk
+
+    elif speaker == "香子":
+        spk = 27
+        return spk
+
+    elif speaker == "真矢":
+        spk = 28
+        return spk
+    
+    elif speaker == "双葉":
+        spk = 29
+        return spk
+    
+    elif speaker == "ミチル":
+        spk = 30
+        return spk
+    
+    elif speaker == "メイファン":
+        spk = 31
+        return spk
+    
+    elif speaker == "やちよ":
+        spk = 32
+        return spk
+   
+    elif speaker == "晶":
+        spk = 33
+        return spk
+    
+    elif speaker == "いちえ":
+        spk = 34
+        return spk
+    
+    elif speaker == "ゆゆ子":
+        spk = 35
+        return spk
+    
+    elif speaker == "塁":
+        spk = 36
+        return spk
+    
+    elif speaker == "珠緒":
+        spk = 37
+        return spk
     
-    def infer(self, text ,language, speaker_id,n_scale= 0.667,n_scale_w = 0.8, l_scale = 1):
-        try:
-            speaker_id = int(self.selection(speaker_id))
+    elif speaker == "あるる":
+        spk = 38
+        return spk
+    
+    elif speaker == "ララフィン":
+        spk = 39
+        return spk
+    
+    elif speaker == "美空":
+        spk = 40
+        return spk
+    
+    elif speaker == "静羽":
+        spk = 41
+        return spk
+    
+    else:
+        return 0
+
+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 = self.get_text(self.sle(language,text))
+            stn_tst = get_text(sle(language,text),hps)
             with torch.no_grad():
-                x_tst = stn_tst.unsqueeze(0).to(self.dev)
-                x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(self.dev)
-                sid = torch.LongTensor([speaker_id]).to(self.dev)
-                audio = self.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()
+                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)
-            return (self.hps.data.sampling_rate, audio)
-        except:
-            self.hps = utils.get_hparams_from_file(f"checkpoints/biaobei/config.json")
-            self.net_g = SynthesizerTrn(
-                len(symbols),
-                self.hps.data.filter_length // 2 + 1,
-                self.hps.train.segment_size // self.hps.data.hop_length,
-                n_speakers=self.hps.data.n_speakers,
-                **self.hps.model).to(self.dev)
-            _ = self.net_g.eval()
-            _ = utils.load_checkpoint(f"checkpoints/biaobei/model.pth", self.net_g)
+                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
 
-    def infer2(self, text ,language, speaker_id,n_scale= 0.667,n_scale_w = 0.8, l_scale = 1):
-        speaker_id = int(self.selection(speaker_id))
-        a = ['【','[','(','（']
-        b = ['】',']',')','）']
-        for i in a:
-            text = text.replace(i,'<')
-        for i in b:
-            text = text.replace(i,'>')
-        final_list = self.extrac(text.replace('“','').replace('”',''))
-        audio_fin = []
-        c = 0
-        t = datetime.timedelta(seconds=0)
-        f1 = open("subtitles.srt",'w',encoding='utf-8')
-        for sentence in final_list:
-            c +=1
-            stn_tst = self.get_text(self.sle(language,sentence))
-            with torch.no_grad():
-                x_tst = stn_tst.unsqueeze(0).to(self.dev)
-                x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(self.dev)
-                sid = torch.LongTensor([speaker_id]).to(self.dev)
-                t1 = time.time()
-                audio = self.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)
-        file_path = "subtitles.srt"
-        return (self.hps.data.sampling_rate, np.concatenate(audio_fin)),file_path
-print("开始部署")
-grVits = VitsGradio()
-grVits.Vits.launch()
+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","ShojoKageki","ShojoKageki-Nijigasaki"]
+    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"]
+        phone_dict = {
+            symbol: i for i, symbol in enumerate(symbols)
+        }
+        checkpoint = models_info[i]["checkpoint"]
+        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)
+    idols = ["c1","c2","高咲侑","歩夢","かすみ","しずく","果林","愛","彼方","せつ菜","璃奈","栞子","エマ","ランジュ","ミア","華恋","まひる","なな","クロディーヌ","ひかり",'純那',"香子","真矢","双葉","ミチル","メイファン","やちよ","晶","いちえ","ゆゆ子","塁","珠緒","あるる","ララフィン","美空","静羽","あるる"]
+    with gr.Blocks() as app:
+        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(
+                                            '<div align="center">'
+                                            f'<img style="width:auto;height:400px;" src="file/image/{name}.png">' 
+                                            '</div>'
+                                        )
+                                    output_UI = gr.outputs.HTML()
+                                with gr.Row():
+                                    with gr.Column(scale=0.85):
+                                        input1 = gr.TextArea(label="Text", value=example,lines = 1)    
+                                    with gr.Column(scale=0.15, min_width=0):
+                                        btnVC = gr.Button("Send")
+                                output1 = gr.Audio(label="采样率22050")
+                                with gr.Accordion(label="Setting(TTS)", open=False):
+                                    input2 = gr.Dropdown(label="Language", choices=lan, value="自动", interactive=True)
+                                    input4 = gr.Slider(minimum=0, maximum=1.0, label="更改噪声比例(noise scale)，以控制情感", value=0.6)
+                                    input5 = gr.Slider(minimum=0, maximum=1.0, label="更改噪声偏差(noise scale w)，以控制音素长短", value=0.668)
+                                    input6 = gr.Slider(minimum=0.1, maximum=10, label="duration", value=1) 
+                                with gr.Accordion(label="Advanced Setting(GPT3.5接口+长句子合成，建议克隆本仓库后运行main.py)", open=False):
+                                    input3 = gr.Checkbox(value=False, label="长句切割(小说合成)")
+                                    output2 = gr.outputs.File(label="字幕文件：subtitles.srt")
+                                    api_input1 = gr.Checkbox(value=False, label="接入chatgpt")
+                                    api_input2 = gr.TextArea(label="api-key",lines=1,value = '见 https://openai.com/blog/openai-api')   
+                                    audio_input1 = gr.Checkbox(value=False, label="修改音频路径(live2d)")
+                                    audio_input2 = gr.TextArea(label="音频路径",lines=1,value = '#参考 D:/app_develop/live2d_whole/2010002/sounds/temp.wav')
+                                    audio_input3 = gr.Dropdown(label="重复生成次数", choices=list(range(101)), value='0', interactive=True) 
+                        btnVC.click(tts_fn, inputs=[api_input1,api_input2,audio_input1,audio_input2,audio_input3,input1,input2,input3,input4,input5,input6], outputs=[output1,output2,output_UI])
+            with gr.Tab("Voice Conversion(弱化版sovits)"):
+                gr.Markdown("""
+                                录制或上传声音，并选择要转换的音色。
+                """)
+                with gr.Column():
+                    record_audio = gr.Audio(label="record your voice", source="microphone")
+                    upload_audio = gr.Audio(label="or upload audio here", source="upload")
+                    source_speaker = gr.Dropdown(choices=idols, value="歩夢", label="source speaker")
+                    target_speaker = gr.Dropdown(choices=idols, value="まひる", label="target speaker")
+                with gr.Column():
+                    message_box = gr.Textbox(label="Message")
+                    converted_audio = gr.Audio(label='converted audio')
+                btn = gr.Button("Convert!")
+                btn.click(vc_fn, inputs=[source_speaker, target_speaker, record_audio, upload_audio],
+                        outputs=[message_box, converted_audio])    
+            with gr.Tab("说明"):
+                gr.Markdown(
+                "### <center> 请不要生成会对个人以及企划造成侵害的内容，自觉遵守相关法律，静止商业使用或让他人产生困扰\n"
+                "<div align='center'>从左到右分别是虹团，少歌中文特化版，以及五校混合版。这三个均为不同的模型，效果也有差异</div>\n"
+                "<div align='center'>因为我会时不时地更新模型，所以会碰到平台抽风问题，大部分情况下一天就能恢复了。</div>\n"
+                '<div align="center"><a>参数说明:这个十分���学，我还没找到最合适的，如果效果不佳可以将噪声比例和噪声偏差调节至0。按照经验，合成日语时也可以将噪声比例调节至0.2-0.3区间，语调会正常一些。duration代表整体语速，1.0大部分情况应该就够了</div>'
+                '<div align="center"><a>建议只在平台上体验最基础的功能，强烈建议将该仓库克隆至本地或者于colab运行 main.py或app.py</div>')
+    app.launch()