Older English dataset question

Common Voice
#1

Apologies if this question has already been answered and I’m not sure where this question goes, but I downloaded an earlier version of the English dataset in June 2018 and it contains ~469h, with the following structure:

/cv-invalid
/cv-other-dev
/cv-other-test
/cv-other-train
/cv-valid-dev
/cv-valid-test
/cv-valid-train

and audio as: sample-000000.mp3, sample-000001.mp3, sample-000002.mp3, sample-000003.mp3, etc. within each subdirectory.

Visiting CommonVoice recently, I see there’s a lot more en data since 2018, but the current version has a different structure and audio file naming convention. Assuming this is a superset of my older version, I’m hoping to use it as test and the larger recent complement as a training set.

To do this, I need to know which audio files overlap in the current version, any chance there is a key for this? Any help would be appreciated.

Looking for Common Voice Corpus English before 2019-02-25 (v1) release
#2

Does your dataset include the file describing each file and the corresponding sentence? It should include an id.

#3

Unfortunately, no client ID in the .csv files in subfolders of the 2018 set. The only columns are:

filename text up_votes down_votes age gender accent duration

#4

Let me check with the team in case someone knows better about the 2018 dataset.

#5

@mhenretty is this something you remember how we generated at the time?

The team suspects IDs might just be continuous numbers but we are not sure.

#6

I’ll keep my fingers crossed someone remembers but I can always use a different training set instead.

Thanks for checking with the team, I really appreciate the data!

#7

@h_caulfield, @nukeador any information regarding the following will be greatly appreciated:

  • Is it possible to download this dataset? It seems that it is no longer available from Common Voice Datasets. Although Common Voice Corpus 1 (2019-02-25) is available for download, this version is different from your version since the file names for Common Voice Corpus 1 (2019-02-25) do not follow sample_#+.mp3 convention of your version.

  • Any luck with finding out overlaps since the file name convention has change? Just curious as to whether this has been resolved.

Lasty, although not related to this topic, I wish the Mozilla Discourse support “Accepted Answer” feature like in stackoverflow so that we can know whether it has been resolved. I do not see such feature mentioned in Who has which moderation powers on Discourse?. I guess Mozilla Discourse is different from Q&A type of community i.e. stackoverflow.