librosa加载CREMA数据集wav报未知格式及NoBackendError错误

对CREMA、RAVDESS、TESS、SAVEE四个语音情感数据集标注情绪标签，将每个音频的存储路径、对应情绪标签整合为DataFrame后导出为csv文件。后续从csv读取音频路径调用librosa.load()接口加载音频时，仅CREMA数据集的wav文件加载失败，其余三个数据集的wav文件均可正常打开。已核查确认文件路径配置正确，当前使用的librosa版本为0.9.1。

报错信息 #

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
File ~\.conda\envs\nhashemi\lib\site-packages\librosa\core\audio.py:155, in load(path, sr, mono, offset, duration, dtype, res_type)
    153 else:
    154     # Otherwise, create the soundfile object
--> 155     context = sf.SoundFile(path)
    157 with context as sf_desc:

File ~\.conda\envs\nhashemi\lib\site-packages\soundfile.py:629, in SoundFile.__init__(self, file, mode, samplerate, channels, subtype, endian, format, closefd)
    627 self._info = _create_info_struct(file, mode, samplerate, channels,
    628                                  format, subtype, endian)
--> 629 self._file = self._open(file, mode_int, closefd)
    630 if set(mode).issuperset('r+') and self.seekable():
    631     # Move write position to 0 (like in Python file objects)

File ~\.conda\envs\nhashemi\lib\site-packages\soundfile.py:1183, in SoundFile._open(self, file, mode_int, closefd)
   1182     raise TypeError("Invalid file: {0!r}".format(self.name))
-> 1183 _error_check(_snd.sf_error(file_ptr),
   1184              "Error opening {0!r}: ".format(self.name))
   1185 if mode_int == _snd.SFM_WRITE:
   1186     # Due to a bug in libsndfile version <= 1.0.25, frames != 0
   1187     # when opening a named pipe in SFM_WRITE mode.
   1188     # See http://github.com/erikd/libsndfile/issues/77.

File ~\.conda\envs\nhashemi\lib\site-packages\soundfile.py:1357, in _error_check(err, prefix)
   1356 err_str = _snd.sf_error_number(err)
-> 1357 raise RuntimeError(prefix + _ffi.string(err_str).decode('utf-8', 'replace'))

RuntimeError: Error opening 'C:/Users/external_dipf/Documents/Dataset/CREMA/AudioWAV/1001_IEO_FEA_HI.wav': File contains data in an unknown format.

During handling of the above exception, another exception occurred:

NoBackendError                            Traceback (most recent call last)
Input In [553], in <cell line: 3>()
      1 emotion='fear'
      2 path = np.array(data_path.Path[data_path.Emotions==emotion])[1]
----> 3 data, sampling_rate = librosa.load(path)
      4 create_waveplot(data, sampling_rate, emotion)
      5 create_spectrogram(data, sampling_rate, emotion)

File ~\.conda\envs\nhashemi\lib\site-packages\librosa\util\decorators.py:88, in deprecate_positional_args.<locals>._inner_deprecate_positional_args.<locals>.inner_f(*args, **kwargs)
     86 extra_args = len(args) - len(all_args)
     87 if extra_args <= 0:
---> 88     return f(*args, **kwargs)
     90 # extra_args > 0
     91 args_msg = [
     92     "{}={}".format(name, arg)
     93     for name, arg in zip(kwonly_args[:extra_args], args[-extra_args:])
     94 ]

File ~\.conda\envs\nhashemi\lib\site-packages\librosa\core\audio.py:174, in load(path, sr, mono, offset, duration, dtype, res_type)
    172 if isinstance(path, (str, pathlib.PurePath)):
    173     warnings.warn("PySoundFile failed. Trying audioread instead.", stacklevel=2)
--> 174     y, sr_native = __audioread_load(path, offset, duration, dtype)
    175 else:
    176     raise (exc)

File ~\.conda\envs\nhashemi\lib\site-packages\librosa\core\audio.py:198, in __audioread_load(path, offset, duration, dtype)
    192 """Load an audio buffer using audioread.
    193 
    194 This loads one block at a time, and then concatenates the results.
    195 """
    197 y = []
--> 198 with audioread.audio_open(path) as input_file:
    199     sr_native = input_file.samplerate
    200     n_channels = input_file.channels

File ~\.conda\envs\nhashemi\lib\site-packages\audioread\__init__.py:116, in audio_open(path, backends)
    113         pass
    115 # All backends failed!
--> 116 raise NoBackendError()

NoBackendError: 

现有实现代码 #

音频情绪标签标注代码 #

CREMA ="C:/Users/external_dipf/Documents/Dataset/CREMA/AudioWAV/"
crema_directory_list = os.listdir(CREMA)

file_emotion = []
file_path = []

for file in crema_directory_list:
    # storing file paths
    file_path.append(CREMA + file)
    # storing file emotions
    part=file.split('_')
    if part[2] == 'SAD':
        file_emotion.append('sad')
    elif part[2] == 'ANG':
        file_emotion.append('angry')
    elif part[2] == 'DIS':
        file_emotion.append('disgust')
    elif part[2] == 'FEA':
        file_emotion.append('fear')
    elif part[2] == 'HAP':
        file_emotion.append('happy')
    elif part[2] == 'NEU':
        file_emotion.append('neutral')
    else:
        file_emotion.append('Unknown')
        
# dataframe for emotion of files
emotion_df = pd.DataFrame(file_emotion, columns=['Emotions'])

# dataframe for path of files.
path_df = pd.DataFrame(file_path, columns=['Path'])
CREMA_df = pd.concat([emotion_df, path_df], axis=1)
CREMA_df.head()

数据集整合与csv存储代码 #

data_path = pd.concat([CREMA_df, RAVDESS_df, TESS_df, SAVEE_df], axis = 0)
data_path.to_csv("data_path.csv",index=False)
data_path.head()

音频加载与可视化代码 #

emotion='fear'
path = np.array(data_path.Path[data_path.Emotions==emotion])[1]
data, sampling_rate = librosa.load(path)
create_waveplot(data, sampling_rate, emotion)
create_spectrogram(data, sampling_rate, emotion)
Audio(path)

报错核心原因有两个：

1. CREMA数据集原生WAV文件采用WAVEFORMATEXTENSIBLE头封装，旧版本libsndfile（soundfile依赖的底层解码库）无法识别该格式，因此抛出「未知数据格式」错误
2. soundfile解码失败后librosa会自动降级调用audioread解码，但当前环境缺少ffmpeg等可用解码后端，因此抛出NoBackendError错误

可按优先级选择以下方案修复：

• 方案1（零代码修改，最快生效）：在当前conda环境安装ffmpeg，为audioread提供解码后端，执行命令conda install -c conda-forge ffmpeg，安装完成后重启Python解释器即可正常加载音频，不需要修改原有路径或业务代码。
• 方案2（彻底解决格式兼容问题）：批量将CREMA数据集音频转码为标准16bit PCM WAV格式，转码后原有路径无需改动，所有音频库均可正常读取。转码参考代码：

from pydub import AudioSegment
import os
from tqdm import tqdm

CREMA_PATH = "C:/Users/external_dipf/Documents/Dataset/CREMA/AudioWAV/"
for file in tqdm(os.listdir(CREMA_PATH)):
    if file.endswith(".wav"):
        full_path = os.path.join(CREMA_PATH, file)
        audio = AudioSegment.from_file(full_path)
        audio.export(full_path, format="wav", parameters=["-acodec", "pcm_s16le"])

• 前置校验：如果单个文件用系统自带播放器也无法打开，说明是压缩包解压不完整导致文件头损坏，重新解压CREMA数据集压缩包或重新下载损坏文件即可。

内容的提问来源于stack exchange，提问作者Niboo