update technical metadata

mdeff / fma

FMA: A Dataset For Music Analysis

https://arxiv.org/abs/1612.01840

MIT License

2.21k stars 432 forks source link

update technical metadata #4

Closed keunwoochoi closed 4 years ago

keunwoochoi commented 7 years ago

As title says. Probably it's not valid?

tae-jun commented 7 years ago

👍 In my case, it's only 1.8KB

mdeff commented 7 years ago

Indeed. Looks like the clipping failed. I have:

fma_full/133/133297.mp3, 4546530 bytes, sha1 d6a6aa0d0a4264e7679aaca3c3cc4c67127143b7
fma_large/133/133297.mp3, 1752 bytes, sha1 19c8a6a03db7df5768cc0a424344cb2441875cb0

mdeff commented 7 years ago

Found the problem. The metadata, scraped from the FMA API, claims the song is 600 seconds long, while ffmpeg -i 133297.mp3 -f null - tells us it's about 01:53.56. The trim_audio() function in creation.py is clipping 30 seconds in the middle of the song, that is from 285s to 315s for this particular file. But that is out of the song, which is 114s long.

It's not the first time I see an error in the reported song length. The solution is to extract the duration from the file with ffmpeg and update the metadata. I should do that at some point.

tae-jun commented 7 years ago

@mdeff You are the best! 👍

mdeff commented 7 years ago

Haha thanks @tae-jun :)

While at this, we should collect the following technical metadata to be included in tracks.csv:

Bit rate
Number of channels
Sample rate
Number of samples
(Duration is then a function of the above two)

I'll probably do that for the final release. If somebody wants to help, you're welcome. :)

chiwanpark commented 7 years ago

I just found same error in fma_small/108/108925.mp3.

mdeff commented 7 years ago

Thanks all for reporting. :) I will fix this problem for the v1 release. Don't hesitate to open issues for other problems if you find some.