Hugging Face's logo

Datasets:
proxectonos
/
Nos_Parlaspeech-GL

Tasks:
Automatic Speech Recognition
Modalities:
Audio
Text
Formats:
parquet
Languages:
Galician
Size:
100K - 1M
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files
xet
Community
2

Potentially Corrupted Audio Files

#2
by gegallego - opened
Discussion
gegallego

Hi,

I believe some of the audio files in the other subcorpus may be corrupted. Below is a list of a few affected files, though there appear to be more:

DSPG_104_27022019_21392.16_21410.13.wav
DSPG_104_27022019_21586.05_21595.95.wav
DSPG_104_27022019_21596.31_21610.83.wav
DSPG_104_27022019_15283.53_15292.08.wav
DSPG_104_27022019_15292.53_15299.94.wav
DSPG_104_27022019_15312.69_15318.27.wav
DSPG_104_27022019_15378.42_15388.05.wav
DSPG_104_27022019_15494.64_15497.94.wav
DSPG_104_27022019_15543.54_15549.27.wav
...

I re-downloaded the corresponding Parquet files multiple times to verify, but it's possible I made a mistake. Could you please help confirm whether these files are indeed corrupted?

Best regards,
Gerard

cmagui
 Proxecto Nós org

Hi Gerard,

Thanks a lot for letting us know! So far, we've only experimented with the clean subcorpus, so it's possible there's some issue in the other subcorpus that we haven't detected yet. We'll take a look and get back to you.

Best regards,

Carmen

cmagui
 Proxecto Nós org

Hi Gerard,

Thanks again for pointing this out! We've looked into the other subcorpus and can confirm that there were indeed corrupted audio files, just as you indicated.

We actually detected and corrected this problem a few months ago. However, when we updated the corpus in Hugging Face, only the clean subcorpus was used, leaving the other subcorpus un-updated. We apologize for this oversight!

In addition to the corrupted audio files, we've also found some segments that are several minutes long. These could potentially cause memory problems during training. We're still investigating the issue with these extra-long segments. In the meantime, if you plan to use this subcorpus, you could filter out these segments to ensure they don't affect your training process.

We'll let you know when we have an update on the long segments. Please let us know if anything else comes up!

Best regards,

Carmen

gegallego

Hi Carmen,

Thanks a lot for checking and confirming the issue. I appreciate the clarification and the follow-up. Please do keep me posted when the fixed version becomes available.

Best regards,
Gerard

Sign up or log in to comment