214 lines
10 KiB
Python
214 lines
10 KiB
Python
import io
|
||
import os
|
||
|
||
# os.system("wget -P cvec/ https://huggingface.co/spaces/innnky/nanami/resolve/main/checkpoint_best_legacy_500.pt")
|
||
import gradio as gr
|
||
import gradio.processing_utils as gr_pu
|
||
import librosa
|
||
import numpy as np
|
||
import soundfile
|
||
from inference.infer_tool import Svc
|
||
import logging
|
||
|
||
import subprocess
|
||
import edge_tts
|
||
import asyncio
|
||
from scipy.io import wavfile
|
||
import librosa
|
||
import torch
|
||
import time
|
||
|
||
logging.getLogger('numba').setLevel(logging.WARNING)
|
||
logging.getLogger('markdown_it').setLevel(logging.WARNING)
|
||
logging.getLogger('urllib3').setLevel(logging.WARNING)
|
||
logging.getLogger('matplotlib').setLevel(logging.WARNING)
|
||
logging.getLogger('multipart').setLevel(logging.WARNING)
|
||
|
||
model = None
|
||
spk = None
|
||
debug = False
|
||
|
||
cuda = []
|
||
if torch.cuda.is_available():
|
||
for i in range(torch.cuda.device_count()):
|
||
device_name = torch.cuda.get_device_properties(i).name
|
||
cuda.append(f"CUDA:{i} {device_name}")
|
||
|
||
def modelAnalysis(model_path,config_path,cluster_model_path,device,enhance):
|
||
global model
|
||
try:
|
||
model = Svc(model_path.name, config_path.name, device=device if device!="Auto" else None, cluster_model_path = cluster_model_path.name if cluster_model_path != None else "",nsf_hifigan_enhance=enhance)
|
||
spks = list(model.spk2id.keys())
|
||
device_name = torch.cuda.get_device_properties(model.dev).name if "cuda" in str(model.dev) else str(model.dev)
|
||
msg = f"成功加载模型到设备{device_name}上\n"
|
||
if cluster_model_path is None:
|
||
msg += "未加载聚类模型\n"
|
||
else:
|
||
msg += f"聚类模型{cluster_model_path.name}加载成功\n"
|
||
msg += "当前模型的可用音色：\n"
|
||
for i in spks:
|
||
msg += i + " "
|
||
return sid.update(choices = spks,value=spks[0]), msg
|
||
except Exception as e:
|
||
raise gr.Error(e)
|
||
|
||
|
||
def modelUnload():
|
||
global model
|
||
if model is None:
|
||
return sid.update(choices = [],value=""),"没有模型需要卸载!"
|
||
else:
|
||
model = None
|
||
torch.cuda.empty_cache()
|
||
return sid.update(choices = [],value=""),"模型卸载完毕!"
|
||
|
||
|
||
def vc_fn(sid, input_audio, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,F0_mean_pooling,enhancer_adaptive_key):
|
||
global model
|
||
try:
|
||
if input_audio is None:
|
||
raise gr.Error("你需要上传音频")
|
||
if model is None:
|
||
raise gr.Error("你需要指定模型")
|
||
sampling_rate, audio = input_audio
|
||
# print(audio.shape,sampling_rate)
|
||
audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
|
||
if len(audio.shape) > 1:
|
||
audio = librosa.to_mono(audio.transpose(1, 0))
|
||
temp_path = "temp.wav"
|
||
soundfile.write(temp_path, audio, sampling_rate, format="wav")
|
||
_audio = model.slice_inference(temp_path, sid, vc_transform, slice_db, cluster_ratio, auto_f0, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,F0_mean_pooling,enhancer_adaptive_key)
|
||
model.clear_empty()
|
||
os.remove(temp_path)
|
||
#构建保存文件的路径，并保存到results文件夹内
|
||
try:
|
||
timestamp = str(int(time.time()))
|
||
filename = sid + "_" + timestamp + ".wav"
|
||
output_file = os.path.join("./results", filename)
|
||
soundfile.write(output_file, _audio, model.target_sample, format="wav")
|
||
return f"推理成功，音频文件保存为results/{filename}", (model.target_sample, _audio)
|
||
except Exception as e:
|
||
raise gr.Error(e)
|
||
except Exception as e:
|
||
raise gr.Error(e)
|
||
|
||
|
||
def tts_func(_text,_rate):
|
||
#使用edge-tts把文字转成音频
|
||
# voice = "zh-CN-XiaoyiNeural"#女性，较高音
|
||
# voice = "zh-CN-YunxiNeural"#男性
|
||
voice = "zh-CN-YunxiNeural"#男性
|
||
output_file = _text[0:10]+".wav"
|
||
# communicate = edge_tts.Communicate(_text, voice)
|
||
# await communicate.save(output_file)
|
||
if _rate>=0:
|
||
ratestr="+{:.0%}".format(_rate)
|
||
elif _rate<0:
|
||
ratestr="{:.0%}".format(_rate)#减号自带
|
||
|
||
p=subprocess.Popen(["edge-tts",
|
||
"--text",_text,
|
||
"--write-media",output_file,
|
||
"--voice",voice,
|
||
"--rate="+ratestr]
|
||
,shell=True,
|
||
stdout=subprocess.PIPE,
|
||
stdin=subprocess.PIPE)
|
||
p.wait()
|
||
return output_file
|
||
|
||
|
||
def vc_fn2(sid, input_audio, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,text2tts,tts_rate,F0_mean_pooling,enhancer_adaptive_key):
|
||
#使用edge-tts把文字转成音频
|
||
output_file=tts_func(text2tts,tts_rate)
|
||
|
||
#调整采样率
|
||
sr2=44100
|
||
wav, sr = librosa.load(output_file)
|
||
wav2 = librosa.resample(wav, orig_sr=sr, target_sr=sr2)
|
||
save_path2= text2tts[0:10]+"_44k"+".wav"
|
||
wavfile.write(save_path2,sr2,
|
||
(wav2 * np.iinfo(np.int16).max).astype(np.int16)
|
||
)
|
||
|
||
#读取音频
|
||
sample_rate, data=gr_pu.audio_from_file(save_path2)
|
||
vc_input=(sample_rate, data)
|
||
|
||
a,b=vc_fn(sid, vc_input, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,F0_mean_pooling,enhancer_adaptive_key)
|
||
os.remove(output_file)
|
||
os.remove(save_path2)
|
||
return a,b
|
||
|
||
|
||
with gr.Blocks(
|
||
theme=gr.themes.Base(
|
||
primary_hue = gr.themes.colors.green,
|
||
font=["Source Sans Pro", "Arial", "sans-serif"],
|
||
font_mono=['JetBrains mono', "Consolas", 'Courier New']
|
||
),
|
||
) as app:
|
||
with gr.Tabs():
|
||
with gr.TabItem("Inference"):
|
||
gr.Markdown(value="""
|
||
So-vits-svc 4.0 推理 webui
|
||
""")
|
||
with gr.Row(variant="panel"):
|
||
with gr.Column():
|
||
gr.Markdown(value="""
|
||
<font size=2> 模型设置</font>
|
||
""")
|
||
model_path = gr.File(label="选择模型文件")
|
||
config_path = gr.File(label="选择配置文件")
|
||
cluster_model_path = gr.File(label="选择聚类模型文件（没有可以不选）")
|
||
device = gr.Dropdown(label="推理设备，默认为自动选择CPU和GPU", choices=["Auto",*cuda,"CPU"], value="Auto")
|
||
enhance = gr.Checkbox(label="是否使用NSF_HIFIGAN增强,该选项对部分训练集少的模型有一定的音质增强效果，但是对训练好的模型有反面效果，默认关闭", value=False)
|
||
with gr.Column():
|
||
gr.Markdown(value="""
|
||
<font size=3>左侧文件全部选择完毕后(全部文件模块显示download)，点击“加载模型”进行解析：</font>
|
||
""")
|
||
model_load_button = gr.Button(value="加载模型", variant="primary")
|
||
model_unload_button = gr.Button(value="卸载模型", variant="primary")
|
||
sid = gr.Dropdown(label="音色（说话人）")
|
||
sid_output = gr.Textbox(label="Output Message")
|
||
|
||
|
||
with gr.Row(variant="panel"):
|
||
with gr.Column():
|
||
gr.Markdown(value="""
|
||
<font size=2> 推理设置</font>
|
||
""")
|
||
auto_f0 = gr.Checkbox(label="自动f0预测，配合聚类模型f0预测效果更好,会导致变调功能失效（仅限转换语音，歌声勾选此项会究极跑调）", value=False)
|
||
F0_mean_pooling = gr.Checkbox(label="是否对F0使用均值滤波器(池化)，对部分哑音有改善。注意，启动该选项会导致推理速度下降，默认关闭", value=False)
|
||
vc_transform = gr.Number(label="变调（整数，可以正负，半音数量，升高八度就是12）", value=0)
|
||
cluster_ratio = gr.Number(label="聚类模型混合比例，0-1之间，0即不启用聚类。使用聚类模型能提升音色相似度，但会导致咬字下降（如果使用建议0.5左右）", value=0)
|
||
slice_db = gr.Number(label="切片阈值", value=-40)
|
||
noise_scale = gr.Number(label="noise_scale 建议不要动，会影响音质，玄学参数", value=0.4)
|
||
with gr.Column():
|
||
pad_seconds = gr.Number(label="推理音频pad秒数，由于未知原因开头结尾会有异响，pad一小段静音段后就不会出现", value=0.5)
|
||
cl_num = gr.Number(label="音频自动切片，0为不切片，单位为秒(s)", value=0)
|
||
lg_num = gr.Number(label="两端音频切片的交叉淡入长度，如果自动切片后出现人声不连贯可调整该数值，如果连贯建议采用默认值0，注意，该设置会影响推理速度，单位为秒/s", value=0)
|
||
lgr_num = gr.Number(label="自动音频切片后，需要舍弃每段切片的头尾。该参数设置交叉长度保留的比例，范围0-1,左开右闭", value=0.75)
|
||
enhancer_adaptive_key = gr.Number(label="使增强器适应更高的音域(单位为半音数)|默认为0", value=0)
|
||
with gr.Tabs():
|
||
with gr.TabItem("音频转音频"):
|
||
vc_input3 = gr.Audio(label="选择音频")
|
||
vc_submit = gr.Button("音频转换", variant="primary")
|
||
with gr.TabItem("文字转音频"):
|
||
text2tts=gr.Textbox(label="在此输入要转译的文字。注意，使用该功能建议打开F0预测，不然会很怪")
|
||
tts_rate = gr.Number(label="tts语速", value=0)
|
||
vc_submit2 = gr.Button("文字转换", variant="primary")
|
||
with gr.Row():
|
||
with gr.Column():
|
||
vc_output1 = gr.Textbox(label="Output Message")
|
||
with gr.Column():
|
||
vc_output2 = gr.Audio(label="Output Audio", interactive=False)
|
||
|
||
vc_submit.click(vc_fn, [sid, vc_input3, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,F0_mean_pooling,enhancer_adaptive_key], [vc_output1, vc_output2])
|
||
vc_submit2.click(vc_fn2, [sid, vc_input3, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,text2tts,tts_rate,F0_mean_pooling,enhancer_adaptive_key], [vc_output1, vc_output2])
|
||
model_load_button.click(modelAnalysis,[model_path,config_path,cluster_model_path,device,enhance],[sid,sid_output])
|
||
model_unload_button.click(modelUnload,[],[sid,sid_output])
|
||
app.launch()
|
||
|
||
|