在音视频领域，把已经发布的混音歌曲或者音频文件逆向分离一直是世界性的课题。音波混合的物理特性导致在没有原始工程文件的情况下，将其还原和分离是一件很有难度的事情。
言及背景音人声分离技术，就不能不提Spleeter，它是一种用于音频源分离（音乐分离）的开源深度学习算法，由Deezer研究团队开发。使用的是一个性能取向的音源分离算法，并且为用户提供了已经预训练好的模型，能够开箱即用，这也是Spleeter泛用性高的原因之一，关于Spleeter，请移步：人工智能AI库Spleeter免费人声和背景音乐分离实践(Python3.10)，这里不再赘述。
MVSEP-MDX23背景音人声分离技术由Demucs研发，Demucs来自Facebook Research团队，它的发源晚于Spleeter，早于MDX-Net，并且经历过4个大版本的迭代，每一代的模型结构都被大改。Demucs的生成质量从v3开始大幅质变，一度领先行业平均水平，v4是现在最强的开源乐器分离单模型，v1和v2的网络模型被用作MDX-net其中的一部分。
本次我们基于MVSEP-MDX23来对音频的背景音和人声进行分离。
本地分离人声和背景音
如果本地离线运行MVSEP-MDX23，首先克隆代码：
git clone https://github.com/jarredou/MVSEP-MDX23-Colab_v2.git 随后进入项目并安装依赖：
cd MVSEP-MDX23-Colab_v2 pip3 install -r requirements.txt 随后直接进推理即可：
python3 inference.py --input_audio test.wav --output_folder ./results/ 这里将test.wav进行人声分离，分离后的文件在results文件夹生成。
注意推理过程中会将分离模型下载到项目的models目录，极其巨大。
同时推理过程相当缓慢。
这里可以添加--single_onnx参数来提高推理速度，但音质上有一定的损失。
如果本地设备具备12G以上的显存，也可以添加--large_gpu参数来提高推理的速度。
如果本地没有N卡或者显存实在捉襟见肘，也可以通过--cpu参数来使用cpu进行推理，但是并不推荐这样做，因为本来就慢，用cpu就更慢了。
令人暖心的是，官方还利用Pyqt写了一个小的gui界面来提高操作友好度：
__author__ = 'Roman Solovyev (ZFTurbo), IPPM RAS' if __name__ == '__main__': import os gpu_use = "0" print('GPU use: {}'.format(gpu_use)) os.environ["CUDA_VISIBLE_DEVICES"] = "{}".format(gpu_use) import time import os import numpy as np from PyQt5.QtCore import * from PyQt5 import QtCore from PyQt5.QtWidgets import * import sys from inference import predict_with_model root = dict() class Worker(QObject): finished = pyqtSignal() progress = pyqtSignal(int) def __init__(self, options): super().__init__() self.options = options def run(self): global root # Here we pass the update_progress (uncalled!) self.options['update_percent_func'] = self.update_progress predict_with_model(self.options) root['button_start'].setDisabled(False) root['button_finish'].setDisabled(True) root['start_proc'] = False self.finished.emit() def update_progress(self, percent): self.progress.emit(percent) class Ui_Dialog(object): def setupUi(self, Dialog): global root Dialog.setObjectName("Settings") Dialog.resize(370, 180) self.checkbox_cpu = QCheckBox("Use CPU instead of GPU?", Dialog) self.checkbox_cpu.move(30, 10) self.checkbox_cpu.resize(320, 40) if root['cpu']: self.checkbox_cpu.setChecked(True) self.checkbox_single_onnx = QCheckBox("Use single ONNX?", Dialog) self.checkbox_single_onnx.move(30, 40) self.checkbox_single_onnx.resize(320, 40) if root['single_onnx']: self.checkbox_single_onnx.setChecked(True) self.pushButton_save = QPushButton(Dialog) self.pushButton_save.setObjectName("pushButton_save") self.pushButton_save.move(30, 120) self.pushButton_save.resize(150, 35) self.pushButton_cancel = QPushButton(Dialog) self.pushButton_cancel.setObjectName("pushButton_cancel") self.pushButton_cancel.move(190, 120) self.pushButton_cancel.resize(150, 35) self.retranslateUi(Dialog) QtCore.QMetaObject.connectSlotsByName(Dialog) self.Dialog = Dialog # connect the two functions self.pushButton_save.clicked.connect(self.return_save) self.pushButton_cancel.clicked.connect(self.return_cancel) def retranslateUi(self, Dialog): _translate = QtCore.QCoreApplication.translate Dialog.setWindowTitle(_translate("Settings", "Settings")) self.pushButton_cancel.setText(_translate("Settings", "Cancel")) self.pushButton_save.setText(_translate("Settings", "Save settings")) def return_save(self): global root # print("save") root['cpu'] = self.checkbox_cpu.isChecked() root['single_onnx'] = self.checkbox_single_onnx.isChecked() self.Dialog.close() def return_cancel(self): global root # print("cancel") self.Dialog.close() class MyWidget(QWidget): def __init__(self): super().__init__() self.initUI() def initUI(self): self.resize(560, 360) self.move(300, 300) self.setWindowTitle('MVSEP music separation model') self.setAcceptDrops(True) def dragEnterEvent(self, event): if event.mimeData().hasUrls(): event.accept() else: event.ignore() def dropEvent(self, event): global root files = [u.toLocalFile() for u in event.mimeData().urls()] txt = '' root['input_files'] = [] for f in files: root['input_files'].append(f) txt += f + 'n' root['input_files_list_text_area'].insertPlainText(txt) root['progress_bar'].setValue(0) def execute_long_task(self): global root if len(root['input_files']) == 0 and 1: QMessageBox.about(root['w'], "Error", "No input files specified!") return root['progress_bar'].show() root['button_start'].setDisabled(True) root['button_finish'].setDisabled(False) root['start_proc'] = True options = { 'input_audio': root['input_files'], 'output_folder': root['output_folder'], 'cpu': root['cpu'], 'single_onnx': root['single_onnx'], 'overlap_large': 0.6, 'overlap_small': 0.5, } self.update_progress(0) self.thread = QThread() self.worker = Worker(options) self.worker.moveToThread(self.thread) self.thread.started.connect(self.worker.run) self.worker.finished.connect(self.thread.quit) self.worker.finished.connect(self.worker.deleteLater) self.thread.finished.connect(self.thread.deleteLater) self.worker.progress.connect(self.update_progress) self.thread.start() def stop_separation(self): global root self.thread.terminate() root['button_start'].setDisabled(False) root['button_finish'].setDisabled(True) root['start_proc'] = False root['progress_bar'].hide() def update_progress(self, progress): global root root['progress_bar'].setValue(progress) def open_settings(self): global root dialog = QDialog() dialog.ui = Ui_Dialog() dialog.ui.setupUi(dialog) dialog.exec_() def dialog_select_input_files(): global root files, _ = QFileDialog.getOpenFileNames( None, "QFileDialog.getOpenFileNames()", "", "All Files (*);;Audio Files (*.wav, *.mp3, *.flac)", ) if files: txt = '' root['input_files'] = [] for f in files: root['input_files'].append(f) txt += f + 'n' root['input_files_list_text_area'].insertPlainText(txt) root['progress_bar'].setValue(0) return files def dialog_select_output_folder(): global root foldername = QFileDialog.getExistingDirectory( None, "Select Directory" ) root['output_folder'] = foldername + '/' root['output_folder_line_edit'].setText(root['output_folder']) return foldername def create_dialog(): global root app = QApplication(sys.argv) w = MyWidget() root['input_files'] = [] root['output_folder'] = os.path.dirname(os.path.abspath(__file__)) + '/results/' root['cpu'] = False root['single_onnx'] = False button_select_input_files = QPushButton(w) button_select_input_files.setText("Input audio files") button_select_input_files.clicked.connect(dialog_select_input_files) button_select_input_files.setFixedHeight(35) button_select_input_files.setFixedWidth(150) button_select_input_files.move(30, 20) input_files_list_text_area = QTextEdit(w) input_files_list_text_area.setReadOnly(True) input_files_list_text_area.setLineWrapMode(QTextEdit.NoWrap) font = input_files_list_text_area.font() font.setFamily("Courier") font.setPointSize(10) input_files_list_text_area.move(30, 60) input_files_list_text_area.resize(500, 100) button_select_output_folder = QPushButton(w) button_select_output_folder.setText("Output folder") button_select_output_folder.setFixedHeight(35) button_select_output_folder.setFixedWidth(150) button_select_output_folder.clicked.connect(dialog_select_output_folder) button_select_output_folder.move(30, 180) output_folder_line_edit = QLineEdit(w) output_folder_line_edit.setReadOnly(True) font = output_folder_line_edit.font() font.setFamily("Courier") font.setPointSize(10) output_folder_line_edit.move(30, 220) output_folder_line_edit.setFixedWidth(500) output_folder_line_edit.setText(root['output_folder']) progress_bar = QProgressBar(w) # progress_bar.move(30, 310) progress_bar.setValue(0) progress_bar.setGeometry(30, 310, 500, 35) progress_bar.setAlignment(QtCore.Qt.AlignCenter) progress_bar.hide() root['progress_bar'] = progress_bar button_start = QPushButton('Start separation', w) button_start.clicked.connect(w.execute_long_task) button_start.setFixedHeight(35) button_start.setFixedWidth(150) button_start.move(30, 270) button_finish = QPushButton('Stop separation', w) button_finish.clicked.connect(w.stop_separation) button_finish.setFixedHeight(35) button_finish.setFixedWidth(150) button_finish.move(200, 270) button_finish.setDisabled(True) button_settings = QPushButton('⚙', w) button_settings.clicked.connect(w.open_settings) button_settings.setFixedHeight(35) button_settings.setFixedWidth(35) button_settings.move(495, 270) button_settings.setDisabled(False) mvsep_link = QLabel(w) mvsep_link.setOpenExternalLinks(True) font = mvsep_link.font() font.setFamily("Courier") font.setPointSize(10) mvsep_link.move(415, 30) mvsep_link.setText('Powered by <a href="https://mvsep.com">MVSep.com</a>') root['w'] = w root['input_files_list_text_area'] = input_files_list_text_area root['output_folder_line_edit'] = output_folder_line_edit root['button_start'] = button_start root['button_finish'] = button_finish root['button_settings'] = button_settings # w.showMaximized() w.show() sys.exit(app.exec_()) if __name__ == '__main__': create_dialog() 效果如下：
界面虽然朴素，但相当实用，Spleeter可没给我们提供这个待遇。
Colab云端分离人声和背景音
托Google的福，我们也可以在Colab云端使用MVSEP-MDX23：
https://colab.research.google.com/github/jarredou/MVSEP-MDX23-Colab_v2/blob/v2.3/MVSep-MDX23-Colab.ipynb#scrollTo=uWX5WOqjU0QC 首先安装MVSEP-MDX23：
#@markdown #Installation #@markdown *Run this cell to install MVSep-MDX23* print('Installing... This will take 1 minute...') %cd /content from google.colab import drive drive.mount('/content/drive') !git clone https://github.com/jarredou/MVSEP-MDX23-Colab_v2.git &> /dev/null %cd /content/MVSEP-MDX23-Colab_v2 !pip install -r requirements.txt &> /dev/null # onnxruntime-gpu nightly fix for cuda12.2 !python -m pip install ort-nightly-gpu --index-url=https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ort-cuda-12-nightly/pypi/simple/ print('Installation done !') 随后编写推理代码：
#@markdown #Separation from pathlib import Path import glob %cd /content/MVSEP-MDX23-Colab_v2 input = '/content/drive/MyDrive' #@param {type:"string"} output_folder = '/content/drive/MyDrive/output' #@param {type:"string"} #@markdown --- #@markdown *Bigshifts=1 to disable that feature* BigShifts = 7 #@param {type:"slider", min:1, max:41, step:1} #@markdown --- overlap_InstVoc = 1 #@param {type:"slider", min:1, max:40, step:1} overlap_VitLarge = 1 #@param {type:"slider", min:1, max:40, step:1} #@markdown --- weight_InstVoc = 8 #@param {type:"slider", min:0, max:10, step:1} weight_VitLarge = 5 #@param {type:"slider", min:0, max:10, step:1} #@markdown --- use_VOCFT = False #@param {type:"boolean"} overlap_VOCFT = 0.1 #@param {type:"slider", min:0, max:0.95, step:0.05} weight_VOCFT = 2 #@param {type:"slider", min:0, max:10, step:1} #@markdown --- vocals_instru_only = True #@param {type:"boolean"} overlap_demucs = 0.6 #@param {type:"slider", min:0, max:0.95, step:0.05} #@markdown --- output_format = 'PCM_16' #@param ["PCM_16", "FLOAT"] if vocals_instru_only: vocals_only = '--vocals_only true' else: vocals_only = '' if use_VOCFT: use_VOCFT = '--use_VOCFT true' else: use_VOCFT = '' if Path(input).is_file(): file_path = input Path(output_folder).mkdir(parents=True, exist_ok=True) !python inference.py --large_gpu --weight_InstVoc {weight_InstVoc} --weight_VOCFT {weight_VOCFT} --weight_VitLarge {weight_VitLarge} --input_audio "{file_path}" --overlap_demucs {overlap_demucs} --overlap_VOCFT {overlap_VOCFT} --overlap_InstVoc {overlap_InstVoc} --overlap_VitLarge {overlap_VitLarge} --output_format {output_format} --BigShifts {BigShifts} --output_folder "{output_folder}" {vocals_only} {use_VOCFT} else: file_paths = sorted([f'"{glob.escape(path)}"' for path in glob.glob(input + "/*")])[:] input_audio_args = ' '.join(file_paths) Path(output_folder).mkdir(parents=True, exist_ok=True) !python inference.py --large_gpu --weight_InstVoc {weight_InstVoc} --weight_VOCFT {weight_VOCFT} --weight_VitLarge {weight_VitLarge} --input_audio {input_audio_args} --overlap_demucs {overlap_demucs} --overlap_VOCFT {overlap_VOCFT} --overlap_InstVoc {int(overlap_InstVoc)} --overlap_VitLarge {int(overlap_VitLarge)} --output_format {output_format} --BigShifts {BigShifts} --output_folder "{output_folder}" {vocals_only} {use_VOCFT} 这里默认使用google云盘的目录，也可以修改为当前服务器的目录地址。
结语
MVSEP-MDX23 和 Spleeter 都是音频人声背景音分离软件，作为用户，我们到底应该怎么选择？
MVSEP-MDX23 基于 Demucs4 和 MDX 神经网络架构，可以将音乐分离成“bass”、“drums”、“vocals”和“other”四个部分。MVSEP-MDX23 在 2023 年的音乐分离挑战中获得了第三名，并且在 MultiSong 数据集上的质量比较中表现出色。它提供了 Python 命令行工具和 GUI 界面，支持 CPU 和 GPU 加速，可以在本地运行。
Spleeter 是由 Deezer 开发的开源音频分离库，它使用深度学习模型将音频分离成不同的音轨，如人声、伴奏等。Spleeter 提供了预训练的模型，可以在命令行或作为 Python 库使用。它的优势在于易用性和灵活性，可以根据需要分离不同数量的音轨。
总的来说，MVSEP-MDX23 在音频分离的性能和精度上表现出色，尤其适合需要高质量音频分离的专业用户。而 Spleeter 则更适合普通用户和开发者，因为它易于使用，并且具有更多的定制选项。
