语音识别实验在训练期间崩溃答案

【问题标题】：Voice recognition experiment crashes during training语音识别实验在训练期间崩溃
【发布时间】：2014-12-09 21:58:16
【问题描述】：

我正在尝试使用 bob.spear 进行扬声器验证实验。

我在训练白化登记者时遇到矛崩溃的问题。

系统规格：

我在 Ubuntu 12.04 上，按照网站上的说明安装了 spear（即首先添加了 bob 1.x 存储库，安装了 bob、bob-dev、sox、sox-dev 软件包以及依赖项，下载了 spear 并使用构建）。

协议：

我已经更改了protocol/mobio/mobile0-female 文件夹中的文件列表，并创建了一个类似的配置。我已将我的文件列表和 dev/eval/norm 文件夹添加到此消息中。

我尝试解决问题的步骤：

重新安装 liblapack（在 ubuntu 中是 liblapack-dev liblapack3g 和我忘记的其他一些包）
重新安装 bob.spear 并在不同的计算机上运行实验
将我的 wav 文件下采样到 16k
确保 for_models.lst、for_probes.lst 中的 dev 和 eval 之间没有重复（尽管这只是为了尝试模仿 mobio 的比赛文件）
确保所有文件列表格式正确

错误日志：

~/bob.spear-1.1.7$ ./bin/spkverif_ivector.py -d config/database/my_database.py -p config/preprocessing/energy.py -f config/features/mfcc_60.py -t config/tools/ivec/ivec_8g_t50.py -z -b ivector_plda  --user-directory results --temp-directory temp
preprocess 23 wave from directory /home/lioruzan/snd_test/ to directory temp/ivector_plda/preprocessed
Input wave file: /home/lioruzan/snd_test/158.wav
No handlers could be found for logger "bob.c++"
After Energy-based VAD there are 142 frames remaining over 231
Input wave file: /home/lioruzan/snd_test/162.wav
After Energy-based VAD there are 168 frames remaining over 306
Input wave file: /home/lioruzan/snd_test/164.wav
After Energy-based VAD there are 208 frames remaining over 293
Input wave file: /home/lioruzan/snd_test/238.wav
After Energy-based VAD there are 88 frames remaining over 140
Input wave file: /home/lioruzan/snd_test/257.wav
After Energy-based VAD there are 236 frames remaining over 335
Input wave file: /home/lioruzan/snd_test/282.wav
After Energy-based VAD there are 145 frames remaining over 218
Input wave file: /home/lioruzan/snd_test/283.wav
After Energy-based VAD there are 148 frames remaining over 268
Input wave file: /home/lioruzan/snd_test/284.wav
After Energy-based VAD there are 292 frames remaining over 383
Input wave file: /home/lioruzan/snd_test/314.wav
After Energy-based VAD there are 231 frames remaining over 248
Input wave file: /home/lioruzan/snd_test/315.wav
After Energy-based VAD there are 165 frames remaining over 270
Input wave file: /home/lioruzan/snd_test/337.wav
After Energy-based VAD there are 188 frames remaining over 210
Input wave file: /home/lioruzan/snd_test/339.wav
After Energy-based VAD there are 80 frames remaining over 180
Input wave file: /home/lioruzan/snd_test/34.wav
After Energy-based VAD there are 68 frames remaining over 125
Input wave file: /home/lioruzan/snd_test/342.wav
After Energy-based VAD there are 82 frames remaining over 213
Input wave file: /home/lioruzan/snd_test/35.wav
After Energy-based VAD there are 140 frames remaining over 315
Input wave file: /home/lioruzan/snd_test/37.wav
After Energy-based VAD there are 83 frames remaining over 145
Input wave file: /home/lioruzan/snd_test/39.wav
After Energy-based VAD there are 127 frames remaining over 258
Input wave file: /home/lioruzan/snd_test/43.wav
After Energy-based VAD there are 131 frames remaining over 245
Input wave file: /home/lioruzan/snd_test/47.wav
After Energy-based VAD there are 62 frames remaining over 108
Input wave file: /home/lioruzan/snd_test/48.wav
After Energy-based VAD there are 128 frames remaining over 158
Input wave file: /home/lioruzan/snd_test/84.wav
After Energy-based VAD there are 167 frames remaining over 304
Input wave file: /home/lioruzan/snd_test/85.wav
After Energy-based VAD there are 144 frames remaining over 165
Input wave file: /home/lioruzan/snd_test/9.wav
After Energy-based VAD there are 236 frames remaining over 325
extract 23 features from wav directory /home/lioruzan/snd_test/ to directory temp/ivector_plda/features
Input wave file : /home/lioruzan/snd_test/158.wav
Input wave file : /home/lioruzan/snd_test/162.wav
Input wave file : /home/lioruzan/snd_test/164.wav
Input wave file : /home/lioruzan/snd_test/238.wav
Warning: something's wrong with the features: temp/ivector_plda/features/238.hdf5
Input wave file : /home/lioruzan/snd_test/257.wav
Input wave file : /home/lioruzan/snd_test/282.wav
Input wave file : /home/lioruzan/snd_test/283.wav
Input wave file : /home/lioruzan/snd_test/284.wav
Input wave file : /home/lioruzan/snd_test/314.wav
Input wave file : /home/lioruzan/snd_test/315.wav
Input wave file : /home/lioruzan/snd_test/337.wav
Input wave file : /home/lioruzan/snd_test/339.wav
Warning: something's wrong with the features: temp/ivector_plda/features/339.hdf5
Input wave file : /home/lioruzan/snd_test/34.wav
Warning: something's wrong with the features: temp/ivector_plda/features/34.hdf5
Input wave file : /home/lioruzan/snd_test/342.wav
Warning: something's wrong with the features: temp/ivector_plda/features/342.hdf5
Input wave file : /home/lioruzan/snd_test/35.wav
Input wave file : /home/lioruzan/snd_test/37.wav
Warning: something's wrong with the features: temp/ivector_plda/features/37.hdf5
Input wave file : /home/lioruzan/snd_test/39.wav
Input wave file : /home/lioruzan/snd_test/43.wav
Input wave file : /home/lioruzan/snd_test/47.wav
Warning: something's wrong with the features: temp/ivector_plda/features/47.hdf5
Input wave file : /home/lioruzan/snd_test/48.wav
Input wave file : /home/lioruzan/snd_test/84.wav
Input wave file : /home/lioruzan/snd_test/85.wav
Input wave file : /home/lioruzan/snd_test/9.wav
Training Projector 'temp/ivector_plda/Projector.hdf5' using 3 training files: 
bob.c++@2014-12-09 22:47:52,247 -- INFO: # KMeansTrainer:
bob.c++@2014-12-09 22:47:52,248 -- INFO: # Iteration 1: 72.976 -> 51.3793
bob.c++@2014-12-09 22:47:52,249 -- INFO: # Iteration 2: 51.3793 -> 50.0718
bob.c++@2014-12-09 22:47:52,249 -- INFO: # Iteration 3: 50.0718 -> 49.4193
bob.c++@2014-12-09 22:47:52,250 -- INFO: # Iteration 4: 49.4193 -> 49.0583
bob.c++@2014-12-09 22:47:52,250 -- INFO: # Iteration 5: 49.0583 -> 48.8668
bob.c++@2014-12-09 22:47:52,251 -- INFO: # Iteration 6: 48.8668 -> 48.7754
bob.c++@2014-12-09 22:47:52,251 -- INFO: # Iteration 7: 48.7754 -> 48.7553
bob.c++@2014-12-09 22:47:52,252 -- INFO: # Iteration 8: 48.7553 -> 48.7513
bob.c++@2014-12-09 22:47:52,252 -- INFO: # EM terminated: likelihood converged
bob.c++@2014-12-09 22:47:52,254 -- INFO: # EMTrainer:
bob.c++@2014-12-09 22:47:52,257 -- INFO: # Iteration 1: -78.1429 -> -77.8048
bob.c++@2014-12-09 22:47:52,259 -- INFO: # Iteration 2: -77.8048 -> -77.4963
bob.c++@2014-12-09 22:47:52,261 -- INFO: # Iteration 3: -77.4963 -> -77.2019
bob.c++@2014-12-09 22:47:52,263 -- INFO: # Iteration 4: -77.2019 -> -77.009
bob.c++@2014-12-09 22:47:52,265 -- INFO: # Iteration 5: -77.009 -> -76.7791
bob.c++@2014-12-09 22:47:52,267 -- INFO: # Iteration 6: -76.7791 -> -76.7064
bob.c++@2014-12-09 22:47:52,269 -- INFO: # Iteration 7: -76.7064 -> -76.6602
bob.c++@2014-12-09 22:47:52,271 -- INFO: # Iteration 8: -76.6602 -> -76.6291
bob.c++@2014-12-09 22:47:52,273 -- INFO: # Iteration 9: -76.6291 -> -76.6005
bob.c++@2014-12-09 22:47:52,275 -- INFO: # Iteration 10: -76.6005 -> -76.598
bob.c++@2014-12-09 22:47:52,275 -- INFO: # EM terminated: likelihood converged
project 23 features from directory temp/ivector_plda/features to directory temp/ivector_plda/projected_ubm using UBM Projector
Training Enroler 'temp/ivector_plda/Enroler.hdf5' using 1 identities: 
bob.c++@2014-12-09 22:47:52,365 -- INFO: # EMTrainer:
bob.c++@2014-12-09 22:47:52,376 -- INFO: # Iteration 1
bob.c++@2014-12-09 22:47:52,382 -- INFO: # Iteration 2
bob.c++@2014-12-09 22:47:52,388 -- INFO: # Iteration 3
bob.c++@2014-12-09 22:47:52,394 -- INFO: # Iteration 4
bob.c++@2014-12-09 22:47:52,400 -- INFO: # Iteration 5
bob.c++@2014-12-09 22:47:52,406 -- INFO: # Iteration 6
bob.c++@2014-12-09 22:47:52,412 -- INFO: # Iteration 7
bob.c++@2014-12-09 22:47:52,418 -- INFO: # Iteration 8
bob.c++@2014-12-09 22:47:52,424 -- INFO: # Iteration 9
bob.c++@2014-12-09 22:47:52,430 -- INFO: # Iteration 10
bob.c++@2014-12-09 22:47:52,437 -- INFO: # Iteration 11
bob.c++@2014-12-09 22:47:52,443 -- INFO: # Iteration 12
bob.c++@2014-12-09 22:47:52,449 -- INFO: # Iteration 13
bob.c++@2014-12-09 22:47:52,455 -- INFO: # Iteration 14
bob.c++@2014-12-09 22:47:52,461 -- INFO: # Iteration 15
bob.c++@2014-12-09 22:47:52,467 -- INFO: # Iteration 16
bob.c++@2014-12-09 22:47:52,473 -- INFO: # Iteration 17
bob.c++@2014-12-09 22:47:52,479 -- INFO: # Iteration 18
bob.c++@2014-12-09 22:47:52,485 -- INFO: # Iteration 19
bob.c++@2014-12-09 22:47:52,491 -- INFO: # Iteration 20
bob.c++@2014-12-09 22:47:52,497 -- INFO: # Iteration 21
bob.c++@2014-12-09 22:47:52,503 -- INFO: # Iteration 22
bob.c++@2014-12-09 22:47:52,509 -- INFO: # Iteration 23
bob.c++@2014-12-09 22:47:52,514 -- INFO: # Iteration 24
bob.c++@2014-12-09 22:47:52,520 -- INFO: # Iteration 25
bob.c++@2014-12-09 22:47:52,520 -- INFO: # EM terminated: maximum number of iterations reached.
project 23 features from directory temp/ivector_plda/projected_ubm to directory temp/ivector_plda/projected_ivector
Training Enroler 'temp/ivector_plda/WhiteEnroler.hdf5' using 1 identities: 
Traceback (most recent call last):
  File "./bin/spkverif_ivector.py", line 21, in <module>
    sys.exit(spear.script.spkverif_ivector.main())
  File "/home/lioruzan/bob.spear-1.1.7/spear/script/spkverif_ivector.py", line 728, in main
    speaker_verify(args)
  File "/home/lioruzan/bob.spear-1.1.7/spear/script/spkverif_ivector.py", line 696, in speaker_verify
    executor.execute_tool_chain()
  File "/home/lioruzan/bob.spear-1.1.7/spear/script/spkverif_ivector.py", line 107, in execute_tool_chain
    self.m_tool_chain.train_whitening_enroler(self.m_tool, dir_type='projected_ivector', force = self.m_args.force)
  File "/home/lioruzan/bob.spear-1.1.7/spear/toolchain/ToolChainIvector.py", line 86, in train_whitening_enroler
    tool.train_whitening_enroler(train_files, str(enroler_file))
  File "/home/lioruzan/bob.spear-1.1.7/spear/tools/IVector.py", line 115, in train_whitening_enroler
    t.train(self.whitening_machine, ivectors_matrix)
RuntimeError: The LAPACK dpotrf function returned a non-zero value.

有人知道可能是什么问题吗？请索取您可能需要的任何信息。

【问题讨论】：

标签： machine-learning signal-processing voice-recognition python-bob

【解决方案1】：

一条消息

 Warning: something's wrong with the features: temp/ivector_plda/features/339.hdf5

表明您的输入文件有问题。由于您没有共享文件，因此很难说到底出了什么问题。可能的问题：

错误的采样率
频道数错误
数据中导致数值溢出的零能量区域

您需要查看产生警告的文件，并思考它们可能有什么问题。

【讨论】：

【解决方案2】：

我认为这是一个有用的信息：信息：# EM 终止：已达到最大迭代次数。文件“/home/lioruzan/bob.spear-1.1.7/spear/script/spkverif_ivector.py”，第 728 行，在 main

也许训练的迭代次数有限，而您尝试做的次数超出了它的处理能力。查看您正在使用的应用程序的规格（文档）。这上面应该写点什么。

或者它应该是它允许多少次迭代的参数，或者它是硬编码的并且你不能使用更多然后是预定义的。

【讨论】：