spleeter 人声分离笔记

      Running from numpy source directory.
      <string>:461: UserWarning: Unrecognized setuptools command, proceeding with generating Cython sources and expanding templates
      Traceback (most recent call last):
        File "B:\ProgramData\miniconda3\envs\py311\Lib\site-packages\pip\_vendor\pyproject_hooks\_in_process\_in_process.py", line 353, in <module>
          main()
        File "B:\ProgramData\miniconda3\envs\py311\Lib\site-packages\pip\_vendor\pyproject_hooks\_in_process\_in_process.py", line 335, in main
          json_out['return_val'] = hook(**hook_input['kwargs'])
                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        File "B:\ProgramData\miniconda3\envs\py311\Lib\site-packages\pip\_vendor\pyproject_hooks\_in_process\_in_process.py", line 149, in prepare_metadata_for_build_wheel
          return hook(metadata_directory, config_settings)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        File "C:\Users\Administrator\AppData\Local\Temp\pip-build-env-mj28xqh3\overlay\Lib\site-packages\setuptools\build_meta.py", line 374, in prepare_metadata_for_build_wheel
          self.run_setup()
        File "C:\Users\Administrator\AppData\Local\Temp\pip-build-env-mj28xqh3\overlay\Lib\site-packages\setuptools\build_meta.py", line 512, in run_setup
          super().run_setup(setup_script=setup_script)
        File "C:\Users\Administrator\AppData\Local\Temp\pip-build-env-mj28xqh3\overlay\Lib\site-packages\setuptools\build_meta.py", line 317, in run_setup
          exec(code, locals())
        File "<string>", line 488, in <module>
        File "<string>", line 465, in setup_package
        File "C:\Users\Administrator\AppData\Local\Temp\pip-install-ehzsep4y\numpy_13571a2d75c64c83ab4a81b86d7d1406\numpy\distutils\__init__.py", line 26, in <module>
          from . import ccompiler
        File "C:\Users\Administrator\AppData\Local\Temp\pip-install-ehzsep4y\numpy_13571a2d75c64c83ab4a81b86d7d1406\numpy\distutils\ccompiler.py", line 111, in <module>
          replace_method(CCompiler, 'find_executables', CCompiler_find_executables)
                         ^^^^^^^^^
      NameError: name 'CCompiler' is not defined. Did you mean: 'ccompiler'?
      [end of output]

# 建议使用新虚拟环境
pip install -U pip setuptools wheel
pip install spleeter  # 默认安装 TensorFlow CPU 版

# 可选：安装 ffmpeg（写 mp3/flac 等需要）
# macOS: brew install ffmpeg
# Ubuntu: sudo apt-get install ffmpeg
# Windows: 安装 ffmpeg 并把 ffmpeg.exe 所在目录加入 PATH

使用 NVIDIA GPU：请改装对应版本的 tensorflow（如 pip install tensorflow==2.12.* 或 tensorflow[and-cuda]，取决于环境）。若只用 CPU，直接上面默认即可。

测试安装是否成功；

python -c "from spleeter.separator import Separator;from spleeter.audio.adapter import AudioAdapter"

方式一：一行命令（最快上手）

# 2 声部（人声 + 伴奏）
spleeter separate -i input.mp3 -p spleeter:2stems -o output_dir

# 4 声部（人声、鼓、贝斯、其他）
spleeter separate -i input.mp3 -p spleeter:4stems -o output_dir

# 5 声部（人声、钢琴、贝斯、鼓、其他）
spleeter separate -i input.wav -p spleeter:5stems -o output_dir

输出会在 output_dir/输入文件名/ 下生成对应的 wav 文件。

方式二：Python 脚本（批量 & 更灵活）

单文件分离（人声/伴奏）



# coding=utf-8
import sys
import os
current_dir = os.path.dirname(os.path.abspath(__file__))
os.chdir(current_dir)

# file: separate_simple.py
from spleeter.separator import Separator
from spleeter.audio.adapter import AudioAdapter
 
def separate_to_files(input_path: str, output_dir: str, stems: int = 2, sample_rate: int = 44100):
    """
    stems: 2 / 4 / 5
    """
    assert stems in (2, 4, 5)
    # 选择模型：2stems/4stems/5stems
    separator = Separator(f'spleeter:{stems}stems')
    audio_loader = AudioAdapter.default()
    # 加载音频为 numpy
    waveform, sr = audio_loader.load(input_path, sample_rate=sample_rate)
    # 执行分离
    prediction = separator.separate(waveform)
    # 将各个轨写入文件（默认写 wav）
    audio_loader.save(f"{output_dir}/vocals.wav", prediction.get('vocals'), sample_rate=sample_rate)
    if stems == 2:
        audio_loader.save(f"{output_dir}/accompaniment.wav", prediction.get('accompaniment'), sample_rate=sample_rate)
    else:
        # 4/5 stems 其他部件可能包括: drums, bass, piano, other
        for k, v in prediction.items():
            if k != 'vocals':
                audio_loader.save(f"{output_dir}/{k}.wav", v, sample_rate=sample_rate)
 
if __name__ == "__main__":

    out_dir=r"out_song"
    os.makedirs(out_dir,exist_ok=True)
    # 示例
    separate_to_files("/nas/lbg/project/tool_down/data_0815/volume_low/1_lang3.mp3", out_dir, stems=2)

批量处理整个文件夹

# file: separate_batch.py
import os
from pathlib import Path
from spleeter.separator import Separator
from spleeter.audio.adapter import AudioAdapter

def batch_separate(input_dir: str, output_root: str, stems: int = 2, sample_rate: int = 44100):
    assert stems in (2, 4, 5)
    separator = Separator(f'spleeter:{stems}stems')   # 初始化一次，重复用
    audio_loader = AudioAdapter.default()

    input_dir = Path(input_dir)
    for p in input_dir.rglob("*"):
        if p.suffix.lower() in {".wav", ".mp3", ".flac", ".m4a", ".ogg"}:
            rel = p.relative_to(input_dir)
            out_dir = Path(output_root) / rel.with_suffix("")  # 每首歌一个文件夹
            out_dir.mkdir(parents=True, exist_ok=True)

            # 加载 & 分离
            waveform, _ = audio_loader.load(str(p), sample_rate=sample_rate)
            prediction = separator.separate(waveform)

            # 保存
            for stem_name, stem_audio in prediction.items():
                audio_loader.save(str(out_dir / f"{stem_name}.wav"), stem_audio, sample_rate=sample_rate)

            # 2 stems 时补一份伴奏
            if stems == 2 and "accompaniment" in prediction:
                pass  # 上面已经保存

if __name__ == "__main__":
    # 示例：把 ./songs 里所有音频分离到 ./separated 下
    batch_separate("./songs", "./separated", stems=2)

指定设备（CPU/GPU）与多线程

from spleeter.separator import Separator # 使用 CPU（默认） sep_cpu = Separator('spleeter:2stems') # 或 Separator('spleeter:4stems') # 使用 GPU（需要正确安装 GPU 版 TensorFlow + CUDA/CuDNN） sep_gpu = Separator('spleeter:2stems', params_descriptor=None) # Spleeter 本身不直接选择设备，设备选择由 TensorFlow 决定： # import tensorflow as tf; tf.config.list_physical_devices('GPU') # 若未检测到 GPU，会自动回落到 CPU。