app.py

import sys, os
sys.path.append('BV2')
import torch
import argparse
import BV2.commons 
import BV2.utils
from BV2.models import Synthesizer
from BV2.text.symbols import symbols
from BV2.text import cleaned_text_to_sequence, get_bert
from BV2.text.cleaner import clean_text
import gradio as gr
import soundfile as sf
from datetime import datetime
import pytz

tz = pytz.timezone('Asia/Shanghai')
net_g = None
models = {
    "Mellowdear": "./BV2/MODELS/adorabledarling.pth",
    "MistyNikki": "./BV2/MODELS/nikki9400.pth",
    "Silverleg": "./BV2/MODELS/J8900.pth",
    "Xelo": "./BV2/MODELS/HER_1100.pth",
    "Rrabbitt": "./BV2/MODELS/rabbit4900.pth",
    "VVV": "./BV2/MODELS/v3.pth",
    "AlluWin": "./BV2/MODELS/AW.pth",
    "Hypnosia": "./BV2/MODELS/hypno.pth",
    "PremJ": "./BV2/MODELS/premj.pth",
    "Umemura": "./BV2/MODELS/take2.pth",
    "ArasakaAI": "./BV2/MODELS/Arasaka.pth",
     "Terra": "./BV2/MODELS/TERRA.pth",
}

def get_text(text, language_str, hps):
    norm_text, phone, tone, word2ph = clean_text(text, language_str)
    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)

    if hps.data.add_blank:
        phone = BV2.commons.intersperse(phone, 0)
        tone = BV2.commons.intersperse(tone, 0)
        language = BV2.commons.intersperse(language, 0)
        for i in range(len(word2ph)):
            word2ph[i] = word2ph[i] * 2
        word2ph[0] += 1
    bert = get_bert(norm_text, word2ph, language_str)
    del word2ph

    assert bert.shape[-1] == len(phone)

    phone = torch.LongTensor(phone)
    tone = torch.LongTensor(tone)
    language = torch.LongTensor(language)

    return bert, phone, tone, language

def infer(text, sdp_ratio, noise_scale, noise_scale_w, length_scale, sid, model_dir):
    global net_g
    bert, phones, tones, lang_ids = get_text(text, "ZH", HPS)
    with torch.no_grad():
        x_tst=phones.to(devicee).unsqueeze(0)
        tones=tones.to(devicee).unsqueeze(0)
        lang_ids=lang_ids.to(devicee).unsqueeze(0)
        bert = bert.to(devicee).unsqueeze(0)
        x_tst_lengths = torch.LongTensor([phones.size(0)]).to(devicee)
        del phones
        speakers = torch.LongTensor([HPS.data.spk2id[sid]]).to(devicee)
        audio = net_g.infer(x_tst, x_tst_lengths, speakers, tones, lang_ids, bert, sdp_ratio=sdp_ratio
                           , noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=length_scale)[0][0,0].data.cpu().float().numpy()
        del x_tst, tones, lang_ids, bert, x_tst_lengths, speakers
        sf.write("tmp.wav", audio, 44100)
        return audio



def convert_wav_to_mp3(wav_file):   
    now = datetime.now(tz).strftime('%m%d%H%M%S')    
    os.makedirs('out', exist_ok=True)  
    output_path_mp3 = os.path.join('out', f"{now}.mp3")

    renamed_input_path = os.path.join('in', f"in.wav")
    os.makedirs('in', exist_ok=True)
    os.rename(wav_file.name, renamed_input_path)
    command = ["ffmpeg", "-i", renamed_input_path, "-acodec", "libmp3lame", "-y", output_path_mp3]
    os.system(" ".join(command))
    return output_path_mp3
    
def tts_generator(text, sdp_ratio, noise_scale, noise_scale_w, length_scale, model):
    global net_g,speakers,tz
    now = datetime.now(tz).strftime('%m-%d %H:%M:%S')
    model_path = models[model]
    net_g, _, _, _ = BV2.utils.load_checkpoint(model_path, net_g, None, skip_optimizer=True)
    print(f'✨{now}-开始生成：{text}')
    try:
        with torch.no_grad():
            audio = infer(text, sdp_ratio=sdp_ratio, noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=length_scale, sid=speaker,model_dir=model)
        with open('tmp.wav', 'rb') as wav_file:
            mp3 = convert_wav_to_mp3(wav_file) 
        return "生成语音成功", (HPS.data.sampling_rate, audio), mp3
    except Exception as e:
        return "生成语音失败：" + str(e), None, None

current_dir = os.path.dirname(os.path.abspath(__file__))
config_path = os.path.join(current_dir, "BV2/configs/config.json")

if __name__ == "__main__":
    HPS = BV2.utils.get_hparams_from_file(config_path)
    devicee = "cuda:0" if torch.cuda.is_available() else "cpu"
   
    net_g = Synthesizer(
        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(devicee)
    _ = net_g.eval()

    speaker_ids = HPS.data.spk2id
    speaker = list(speaker_ids.keys())[0]
    theme='remilia/Ghostly'
    
    with gr.Blocks(theme=theme) as app:
        with gr.Column():
            with gr.Column():
                gr.HTML('''<br><br>
    <p style="margin-bottom: 10px; font-size: 120%">
   Use <b>English</b> to generate, please go to this <a href="https://huggingface.co/spaces/Ailyth/Multi-voice-TTS-GPT-SoVITS" target="_blank">SPACE</a>
   </p>
    <p style="margin-bottom: 10px; font-size: 110%">
   <b>日本語</b>で生成するために、<a href="https://huggingface.co/spaces/Ailyth/Multi-voice-TTS-GPT-SoVITS" target="_blank">こちら</a>へ進んでください。
    </p>''')
                gr.HTML('''
                <hr>
                <p style="margin-bottom: 10px; font-size: 130%"><strong>以下仅供测试用，质量参差</strong>Only read Chinese</p>
                <p>
                模型训练以及推理基于开源项目<a href="https://github.com/fishaudio/Bert-VITS2">Bert-VITS2</a>
                （具体使用的是9月份的版本，可能后续项目效果更好，请自行尝试训练）</p>
                ''')
                text = gr.TextArea(label="输入需要生成语音的文字(标点也会影响语气)", placeholder="输入文字",
                                value="今天拿白金了吗",
                                info="使用huggingface的免费CPU进行推理，因此速度不快，一次性不要输入超过500汉字。字数越多，生成速度越慢，请耐心等待，只会说中文。",
                                  )
                model = gr.Radio(choices=list(models.keys()), value=list(models.keys())[0], label='声音模型')
                with gr.Accordion(label="展开设置生成参数", open=False):
                    sdp_ratio = gr.Slider(minimum=0, maximum=1, value=0.2, step=0.01, label='SDP/DP混合比',info='可控制一定程度的语调变化')
                    noise_scale = gr.Slider(minimum=0.1, maximum=1.5, value=0.5, step=0.01, label='感情变化')
                    noise_scale_w = gr.Slider(minimum=0.1, maximum=1.4, value=0.9, step=0.01, label='音节长度')
                    length_scale = gr.Slider(minimum=0.1, maximum=2, value=1, step=0.01, label='生成语音总长度',info='数值越大，语速越慢')
                btn = gr.Button("✨生成", variant="primary")
            with gr.Column():
                audio_output = gr.Audio(label="试听")
                MP3_output = gr.File(label="💾下载")
                text_output = gr.Textbox(label="调试信息")
                gr.Markdown("""
                
                """)
        btn.click(
                tts_generator,
                inputs=[text, sdp_ratio, noise_scale, noise_scale_w, length_scale, model],
                outputs=[text_output, audio_output,MP3_output]
                )
        gr.HTML('''<div align=center><img id="visitor-badge" alt="visitor badge" src="https://visitor-badge.laobi.icu/badge?page_id=Ailyth/DLMP99" /></div>''')
    app.launch(share=True)