import ONNXVITS_models import utils from text import text_to_sequence import torch import commons def get_text(text, hps): text_norm = text_to_sequence(text, hps.symbols, hps.data.text_cleaners) if hps.data.add_blank: text_norm = commons.intersperse(text_norm, 0) text_norm = torch.LongTensor(text_norm) return text_norm hps = utils.get_hparams_from_file("lovelive/config.json") symbols = hps.symbols net_g = ONNXVITS_models.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) _ = net_g.eval() _ = utils.load_checkpoint("lovelive/G_525000.pth", net_g) text1 = get_text("[JA]ありがとうございます。[JA]", hps) stn_tst = text1 with torch.no_grad(): x_tst = stn_tst.unsqueeze(0) x_tst_lengths = torch.LongTensor([stn_tst.size(0)]) sid = torch.tensor([0]) o = net_g(x_tst, x_tst_lengths, sid=sid, noise_scale=.667, noise_scale_w=0.8, length_scale=1) ''' import romajitable import re import numpy as np import logging logging.getLogger('numba').setLevel(logging.WARNING) import IPython.display as ipd import torch import commons import utils import ONNXVITS_infer from text.symbols import symbols from text import text_to_sequence import gradio as gr import time def get_text(text, hps): text_norm = text_to_sequence(text, symbols, hps.data.text_cleaners) if hps.data.add_blank: text_norm = commons.intersperse(text_norm, 0) text_norm = torch.LongTensor(text_norm) return text_norm 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 elif speaker == "ミア": spk = 12 return spk elif speaker == "三色绘恋1": spk = 13 return spk elif speaker == "三色绘恋2": spk = 15 elif speaker == "派蒙": spk = 16 return spk 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 sle(language,tts_input0): if language == "中文": tts_input1 = "[ZH]" + tts_input0.replace('\n','。').replace(' ',',') + "[ZH]" return tts_input1 if language == "自动": tts_input1 = f"[JA]{tts_input0}[JA]" if is_japanese(tts_input0) else f"[ZH]{tts_input0}[ZH]" return tts_input1 elif language == "日文": tts_input1 = "[JA]" + tts_input0.replace('\n','。').replace(' ',',') + "[JA]" return tts_input1 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(' ','') 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 final_list.append(i) final_list = [x for x in final_list if x != ''] print(final_list) return final_list def infer(language,text,speaker_id, n_scale= 0.667,n_scale_w = 0.8, l_scale = 1 ): speaker_id = int(selection(speaker_id)) 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 for sentence in final_list: c +=1 try: stn_tst = get_text(sle(language,sentence), hps_ms) 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_ms.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) audio_fin.append(audio) except: print('存在非法字符') return (hps_ms.data.sampling_rate, np.concatenate(audio_fin)) lan = ["中文","日文","自动"] idols = ["高咲侑","歩夢","かすみ","しずく","果林","愛","せつ菜","璃奈","栞子","エマ","ランジュ","ミア","派蒙"] dev = torch.device("cpu") hps_ms = utils.get_hparams_from_file("lovelive/config.json") net_g_ms = ONNXVITS_infer.SynthesizerTrn( len(symbols), hps_ms.data.filter_length // 2 + 1, hps_ms.train.segment_size // hps_ms.data.hop_length, n_speakers=hps_ms.data.n_speakers, **hps_ms.model) _ = net_g_ms.eval() _ = utils.load_checkpoint("lovelive/G_525000.pth", net_g_ms) app = gr.Blocks() with app: with gr.Tabs(): with gr.TabItem("虹团vits模型，现可按句分割实现长文本合成，onnx导出后存在质量损失，建议本地运行vits模型"): tts_input1 = gr.TextArea(label="去标贝新模型，老版本在lovelive文件夹中", value="数千怀言者已经为你集结，列队在通往主舰桥的过道上。他们歌唱着你们名字，高声呼喊，以一种原始的、咆哮般的合唱作为对你的致敬。你从他们中间走过，一边点头，一边接受他们的赞美，你沉溺其中，几乎被他们巨大的音量所震撼。\n他们之中没有一个胆敢直视你。没有一个能够承受。你对他们超人类的眼睛来说都太过光辉。从他们中间走过时，你巨大的影子从他们身上掠过，他们立时将目光挪开，眼含泪水，吟诵你的大名时甚至不敢看你一眼。他们的吟唱中含有愤怒。几乎是疯狂的绝望。那感觉就好像他们害怕停下来，害怕自己会喘息停顿，好像尖叫出你的名字是唯一能让他们活着的事情。\n或许确实如此。作为对他们崇拜的回应，你谦虚地抬抬手，随后走进主舰桥。\nI In a word, Horus is a joker.") language = gr.Dropdown(label="选择语言，目前勉强可以做到自动识别",choices=lan, value="自动", interactive=True) para_input1 = gr.Slider(minimum= 0.01,maximum=1.0,label="更改噪声比例，以控制情感", value=0.667) para_input2 = gr.Slider(minimum= 0.01,maximum=1.0,label="更改噪声偏差，以控制音素长短", value=0.7) para_input3 = gr.Slider(minimum= 0.1,maximum=10,label="更改时间比例", value=1) tts_submit = gr.Button("Generate", variant="primary") speaker1 = gr.Dropdown(label="选择说话人",choices=idols, value="歩夢", interactive=True) tts_output2 = gr.Audio(label="Output") tts_submit.click(infer, [language,tts_input1,speaker1,para_input1,para_input2,para_input3], [tts_output2]) #app.launch(share=True) app.launch() '''