运行PyTorch音频预处理教程代码遇LibsndfileError求助

问题场景 #

学习PyTorch音频预处理教程时，运行以下代码获取远程WAV文件的元数据时抛出错误：

print("Source:", SAMPLE_WAV_URL)
with requests.get(SAMPLE_WAV_URL, stream=True) as response:
    metadata = torchaudio.info(response.raw)
print(metadata)

错误输出 #

Source: https://pytorch-tutorial-assets.s3.amazonaws.com/steam-train-whistle-daniel_simon.wav
忽略cffi回调函数<function SoundFile._init_virtual_io.<locals>.vio_get_filelen at 0x000002487DE62C10>的异常:
回溯(最近的调用最后):
  文件 "C:\Users\me\anaconda3\envs\myenv\lib\site-packages\soundfile.py", line 1228, in vio_get_filelen
    file.seek(0, SEEK_END)
io.UnsupportedOperation: seek
---------------------------------------------------------------------------
LibsndfileError                           Traceback (most recent call last)
输入 In [5], in <cell line: 2>()
      1 print("Source:", SAMPLE_WAV_URL)
      2 with requests.get(SAMPLE_WAV_URL, stream=True) as response:
----> 3     metadata = torchaudio.info(response.raw)
      4 print(metadata)

文件 ~\anaconda3\envs\myenv\lib\site-packages\torchaudio\backend\soundfile_backend.py:103, in info(filepath, format)
     84 @_mod_utils.requires_soundfile()
     85 def info(filepath: str, format: Optional[str] = None) -> AudioMetaData:
     86     """获取音频文件的信号信息。
     87 
     88     注意:
   (...)
    101 
    102     """
--> 103     sinfo = soundfile.info(filepath)
    104     return AudioMetaData(
    105         sinfo.samplerate,
    106         sinfo.frames,
   (...)
    109         encoding=_get_encoding(sinfo.format, sinfo.subtype),
    110     )

文件 ~\anaconda3\envs\myenv\lib\site-packages\soundfile.py:464, in info(file, verbose)
    456 def info(file, verbose=False):
    457     """返回包含SoundFile信息的对象。
    458 
    459     参数:
   (...)
    462         是否打印额外信息。
    463     """
--> 464     return _SoundFileInfo(file, verbose)

文件 ~\anaconda3\envs\myenv\lib\site-packages\soundfile.py:409, in _SoundFileInfo.__init__(self, file, verbose)
    407 def __init__(self, file, verbose):
    408     self.verbose = verbose
--> 409     with SoundFile(file) as f:
    410         self.name = f.name
    411         self.samplerate = f.samplerate

文件 ~\anaconda3\envs\myenv\lib\site-packages\soundfile.py:655, in SoundFile.__init__(self, file, mode, samplerate, channels, subtype, endian, format, closefd)
    652 self._mode = mode
    653 self._info = _create_info_struct(file, mode, samplerate, channels,
    654                                  format, subtype, endian)
--> 655 self._file = self._open(file, mode_int, closefd)
    656 如果set(mode).issuperset('r+')且self.seekable():
    657     # 像Python文件对象一样将写入位置移到0
    658     self.seek(0)

文件 ~\anaconda3\envs\myenv\lib\site-packages\soundfile.py:1213, in SoundFile._open(self, file, mode_int, closefd)
   1210 如果file_ptr == _ffi.NULL:
   1211     # 获取实际错误代码
   1212     err = _snd.sf_error(file_ptr)
--> 1213     raise LibsndfileError(err, prefix="打开{0!r}时出错: ".format(self.name))
   1214 如果mode_int == _snd.SFM_WRITE:
   1215     # 由于libsndfile版本<=1.0.25的bug，以SFM_WRITE模式打开命名管道时frames != 0
   1216     # 参见http://github.com/erikd/libsndfile/issues/77。
   1217     self._info.frames = 0

最终错误提示：

LibsndfileError: 打开<urllib3.response.HTTPResponse object at 0x000002487DDD02B0>时出错: WAV文件错误。找不到'data'块标记。

问题原因 #

• requests.get(..., stream=True)返回的response.raw是单向流式HTTP响应对象，不支持随机访问（seek操作）。
• torchaudio底层依赖的libsndfile库在读取WAV元数据时，需要通过seek跳转文件指针来定位WAV文件的"data"块，流式对象无法满足这个需求，导致解析失败。

解决方法 #

方法1：先下载文件到本地再读取 #

适合处理大音频文件，避免占用过多内存：

import requests
import torchaudio
import os

SAMPLE_WAV_URL = "https://pytorch-tutorial-assets.s3.amazonaws.com/steam-train-whistle-daniel_simon.wav"
local_file = "steam_train.wav"

# 下载文件到本地
with requests.get(SAMPLE_WAV_URL) as resp:
    with open(local_file, "wb") as f:
        f.write(resp.content)

# 读取本地文件的元数据
metadata = torchaudio.info(local_file)
print(metadata)

方法2：将响应内容转为支持seek的内存对象 #

适合小音频文件，直接在内存中处理：

import requests
import torchaudio
from io import BytesIO

SAMPLE_WAV_URL = "https://pytorch-tutorial-assets.s3.amazonaws.com/steam-train-whistle-daniel_simon.wav"

with requests.get(SAMPLE_WAV_URL) as resp:
    # 将响应内容转为BytesIO，支持seek操作
    audio_buffer = BytesIO(resp.content)
    metadata = torchaudio.info(audio_buffer)
print(metadata)

内容的提问来源于stack exchange，提问作者Imago