Datasets:

ArXiv:
License:

A question regarding the yodas dataset size

#4
by snakers4 - opened

Hi,

Recently found this dataset and played a little bit with the ru000 subset.

It states that this subset has 4968.37 hours of audio, and indeed it downloads about 500GB worth of wav-files. Back-of-the envelope calculation (int16 wav files, 2 bytes per each second of audio) gives me about 533GB for the chunk, which is very close.

But if I successfully process this dataset chunk using the provided utils, I see only 500 text chunks (i.e. 500 audios?) with 127,675 utterances, which yield about 250-300 hours tops.

Am I doing something wrong, or is it intended? Also why use wav, something like ogg/opus or ogg/vorbis would have the same quality and take ~10x less space.

ESPnet org

hi, thanks for playing with it. I will take a look at this subset and get back to you.

ESPnet org

hi, there was an issue related to uncompressing tar.gz. It should be fixed in the latest version. Can you try updating and counting again? thanks!

Sign up or log in to comment