# Agung Wijaya - WebUI 2023 - Gradio
# file app.py
# Import
import os
import psutil
import shutil
import numpy as np
import gradio as gr
import subprocess
from pathlib import Path
import ffmpeg
import json
import re
import time
import random
import torch
import librosa
import util
from config import device
from infer_pack.models import (
SynthesizerTrnMs256NSFsid,
SynthesizerTrnMs256NSFsid_nono
)
from vc_infer_pipeline import VC
from typing import Union
from os import path, getenv
from datetime import datetime
from scipy.io.wavfile import write
from pydub import AudioSegment
title_markdown = ("""
""")
title_markdown2 = ("""
""")
# Reference: https://huggingface.co/spaces/zomehwh/rvc-models/blob/main/app.py#L21 # noqa
in_hf_space = getenv('SYSTEM') == 'spaces'
# Set High Quality (.wav) or not (.mp3)
high_quality = True
# Read config.json
config_json = json.loads(open("config.json").read())
# Load hubert model
hubert_model = util.load_hubert_model(device, 'hubert_base.pt')
hubert_model.eval()
# Load models
loaded_models = []
for model_name in config_json.get('models'):
print(f'Loading model: {model_name}')
# Load model info
model_info = json.load(
open(path.join('model', model_name, 'config.json'), 'r')
)
# Load RVC checkpoint
cpt = torch.load(
path.join('model', model_name, model_info['model']),
map_location='cpu'
)
tgt_sr = cpt['config'][-1]
cpt['config'][-3] = cpt['weight']['emb_g.weight'].shape[0] # n_spk
if_f0 = cpt.get('f0', 1)
net_g: Union[SynthesizerTrnMs256NSFsid, SynthesizerTrnMs256NSFsid_nono]
if if_f0 == 1:
net_g = SynthesizerTrnMs256NSFsid(
*cpt['config'],
is_half=util.is_half(device)
)
else:
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt['config'])
del net_g.enc_q
# According to original code, this thing seems necessary.
print(net_g.load_state_dict(cpt['weight'], strict=False))
net_g.eval().to(device)
net_g = net_g.half() if util.is_half(device) else net_g.float()
vc = VC(tgt_sr, device, util.is_half(device))
loaded_models.append(dict(
name=model_name,
metadata=model_info,
vc=vc,
net_g=net_g,
if_f0=if_f0,
target_sr=tgt_sr
))
print(f'Models loaded: {len(loaded_models)}')
# Command line test
def command_line_test():
command = "df -h /home/user/app"
process = subprocess.run(command.split(), stdout=subprocess.PIPE)
result = process.stdout.decode()
return gr.HTML(value=result)
def mix(audio1, audio2):
sound1 = AudioSegment.from_file(audio1)
sound2 = AudioSegment.from_file(audio2)
length = len(sound1)
mixed = sound1[:length].overlay(sound2)
mixed.export("song.wav", format="wav")
return "song.wav"
# Function YouTube Downloader Audio
def youtube_downloader(
video_identifier,
start_time,
end_time,
output_filename="track.wav",
num_attempts=5,
url_base="",
quiet=False,
force=True,
):
output_path = Path(output_filename)
if output_path.exists():
if not force:
return output_path
else:
output_path.unlink()
quiet = "--quiet --no-warnings" if quiet else ""
command = f"""
yt-dlp {quiet} -x --audio-format wav -f bestaudio -o "{output_filename}" --download-sections "*{start_time}-{end_time}" "{url_base}{video_identifier}" # noqa: E501
""".strip()
attempts = 0
while True:
try:
_ = subprocess.check_output(command, shell=True, stderr=subprocess.STDOUT)
except subprocess.CalledProcessError:
attempts += 1
if attempts == num_attempts:
return None
else:
break
if output_path.exists():
return output_path
else:
return None
# Function Audio Separated
def audio_separated(audio_input, progress=gr.Progress()):
# start progress
progress(progress=0, desc="Starting...")
time.sleep(0.1)
# check file input
if audio_input is None:
# show progress
for i in progress.tqdm(range(100), desc="Please wait..."):
time.sleep(0.01)
return (None, None, 'Please input audio.')
# create filename
filename = str(random.randint(10000,99999))+datetime.now().strftime("%d%m%Y%H%M%S")
# progress
progress(progress=0.10, desc="Please wait...")
# make dir output
os.makedirs("output", exist_ok=True)
# progress
progress(progress=0.20, desc="Please wait...")
# write
if high_quality:
write(filename+".wav", audio_input[0], audio_input[1])
else:
write(filename+".mp3", audio_input[0], audio_input[1])
# progress
progress(progress=0.50, desc="Please wait...")
# demucs process
if high_quality:
command_demucs = "python3 -m demucs --two-stems=vocals -d cpu "+filename+".wav -o output"
else:
command_demucs = "python3 -m demucs --two-stems=vocals --mp3 --mp3-bitrate 128 -d cpu "+filename+".mp3 -o output"
os.system(command_demucs)
# progress
progress(progress=0.70, desc="Please wait...")
# remove file audio
if high_quality:
command_delete = "rm -v ./"+filename+".wav"
else:
command_delete = "rm -v ./"+filename+".mp3"
os.system(command_delete)
# progress
progress(progress=0.80, desc="Please wait...")
# progress
for i in progress.tqdm(range(80,100), desc="Please wait..."):
time.sleep(0.1)
if high_quality:
return "./output/htdemucs/"+filename+"/vocals.wav","./output/htdemucs/"+filename+"/no_vocals.wav","Successfully..."
else:
return "./output/htdemucs/"+filename+"/vocals.mp3","./output/htdemucs/"+filename+"/no_vocals.mp3","Successfully..."
# Function Voice Changer
def voice_changer(audio_input, model_index, pitch_adjust, f0_method, feat_ratio, progress=gr.Progress()):
# start progress
progress(progress=0, desc="Starting...")
time.sleep(1)
# check file input
if audio_input is None:
# progress
for i in progress.tqdm(range(100), desc="Please wait..."):
time.sleep(0.1)
return (None, 'Please input audio.')
# check model input
if model_index is None:
# progress
for i in progress.tqdm(range(100), desc="Please wait..."):
time.sleep(0.1)
return (None, 'Please select a model.')
model = loaded_models[model_index]
# Reference: so-vits
(audio_samp, audio_npy) = audio_input
# progress
progress(progress=0.10, desc="Please wait...")
# https://huggingface.co/spaces/zomehwh/rvc-models/blob/main/app.py#L49
if (audio_npy.shape[0] / audio_samp) > 60 and in_hf_space:
# progress
for i in progress.tqdm(range(10,100), desc="Please wait..."):
time.sleep(0.1)
return (None, 'Input audio is longer than 60 secs.')
# Bloody hell: https://stackoverflow.com/questions/26921836/
if audio_npy.dtype != np.float32: # :thonk:
audio_npy = (
audio_npy / np.iinfo(audio_npy.dtype).max
).astype(np.float32)
# progress
progress(progress=0.30, desc="Please wait...")
if len(audio_npy.shape) > 1:
audio_npy = librosa.to_mono(audio_npy.transpose(1, 0))
# progress
progress(progress=0.40, desc="Please wait...")
if audio_samp != 16000:
audio_npy = librosa.resample(
audio_npy,
orig_sr=audio_samp,
target_sr=16000
)
# progress
progress(progress=0.50, desc="Please wait...")
pitch_int = int(pitch_adjust)
times = [0, 0, 0]
output_audio = model['vc'].pipeline(
hubert_model,
model['net_g'],
model['metadata'].get('speaker_id', 0),
audio_npy,
times,
pitch_int,
f0_method,
path.join('model', model['name'], model['metadata']['feat_index']),
path.join('model', model['name'], model['metadata']['feat_npy']),
feat_ratio,
model['if_f0']
)
# progress
progress(progress=0.80, desc="Please wait...")
print(f'npy: {times[0]}s, f0: {times[1]}s, infer: {times[2]}s')
# progress
for i in progress.tqdm(range(80,100), desc="Please wait..."):
time.sleep(0.1)
return ((model['target_sr'], output_audio), 'Successfully...')
# Function Text to Voice
def text_to_voice(text_input, model_index):
# start progress
progress(progress=0, desc="Starting...")
time.sleep(1)
# check text input
if text_input is None:
# progress
for i in progress.tqdm(range(2,100), desc="Please wait..."):
time.sleep(0.1)
return (None, 'Please write text.')
# check model input
if model_index is None:
# progress
for i in progress.tqdm(range(2,100), desc="Please wait..."):
time.sleep(0.1)
return (None, 'Please select a model.')
# progress
for i in progress.tqdm(range(2,100), desc="Please wait..."):
time.sleep(0.1)
return None, "Sorry, you can't use it yet because this program is being developed!"
# Themes
theme = gr.themes.Base()
# CSS
css = "footer {visibility: hidden}"
# Blocks
with gr.Blocks(theme=theme, css=css) as App:
# Header
gr.HTML(""
"🥳🎶🎡 - AI歌手,RVC歌声转换
"
"")
gr.Markdown("### 🦄 - 能够自动提取视频中的声音,并去除背景音;Powered by [RVC-Project](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI)")
gr.Markdown("### 更多精彩应用,敬请关注[滔滔AI](http://www.talktalkai.com);滔滔AI,为爱滔滔!💕")
# Information
with gr.Accordion("🎙️您的AI歌手Stefanie:"):
gr.Markdown(title_markdown)
with gr.Accordion("🎤您的AI歌手Vae:"):
gr.Markdown(title_markdown2)
# Tab YouTube Downloader
with gr.Tab("🤗 - B站视频提取声音"):
with gr.Row():
with gr.Column():
ydl_url_input = gr.Textbox(label="B站视频网址(请填写相应的BV号)", value = "https://www.bilibili.com/video/BV...")
start = gr.Number(value=0, label="起始时间 (秒)")
end = gr.Number(value=15, label="结束时间 (秒)")
ydl_url_submit = gr.Button("提取声音文件吧", variant="primary")
as_audio_submit = gr.Button("去除背景音吧", variant="primary")
with gr.Column():
ydl_audio_output = gr.Audio(label="Audio from Bilibili")
as_audio_input = ydl_audio_output
as_audio_vocals = gr.Audio(label="Vocal only")
as_audio_no_vocals = gr.Audio(label="Music only", type="filepath")
as_audio_message = gr.Textbox(label="Message", visible=False)
ydl_url_submit.click(fn=youtube_downloader, inputs=[ydl_url_input, start, end], outputs=[ydl_audio_output])
as_audio_submit.click(fn=audio_separated, inputs=[as_audio_input], outputs=[as_audio_vocals, as_audio_no_vocals, as_audio_message], show_progress=True, queue=True)
# Tab Voice Changer
with gr.Tab("🎶 - 歌声转换"):
with gr.Row():
with gr.Column():
vc_audio_input = as_audio_vocals
vc_model_index = gr.Dropdown(
[
'%s' % (
m['metadata'].get('name')
)
for m in loaded_models
],
label='Models',
type='index'
)
vc_pitch_adjust = gr.Slider(label='Pitch', minimum=-24, maximum=24, step=1, value=0)
vc_f0_method = gr.Radio(label='F0 methods', choices=['pm', 'harvest'], value='pm', interactive=True)
vc_feat_ratio = gr.Slider(label='Feature ratio', minimum=0, maximum=1, step=0.1, value=0.6)
vc_audio_submit = gr.Button("进行歌声转换吧!", variant="primary")
full_song = gr.Button("加入歌曲伴奏吧!", variant="primary")
with gr.Column():
vc_audio_output = gr.Audio(label="Result audio", type="filepath")
vc_audio_message = gr.Textbox(label="Message")
new_song = gr.Audio(label="Full song", type="filepath")
vc_audio_submit.click(fn=voice_changer, inputs=[vc_audio_input, vc_model_index, vc_pitch_adjust, vc_f0_method, vc_feat_ratio], outputs=[vc_audio_output, vc_audio_message], show_progress=True, queue=True)
full_song.click(fn=mix, inputs=[vc_audio_output, as_audio_no_vocals], outputs=[new_song])
gr.Markdown("### 注意❗:请不要生成会对个人以及组织造成侵害的内容,此程序仅供科研、学习及个人娱乐使用。")
gr.HTML('''
''')
# Launch
App.queue(concurrency_count=1, max_size=20).launch(server_name="0.0.0.0", server_port=7860)
# Enjoy