audio-captioning / clotho-dataset Goto Github PK

I am curious if there was any reason or you did not care about audio files format.

I am considering creating a large scale audio dataset by training a small model that later I can use to filter millions of samples and file format seems sensible to be as compressed as possible to save storage space.

Hi, thanks for the great work!
I'm new to the dataset; and I found out there are overlapping entries between split sets as below.

Between valid and eval sets:

• sound_id 66304. In valid set, the segment is [1607168, 2436248]. In eval, it's [1597440, 2865471].
  • file name: 01 A pug struggles to breathe 1_14_2008.wav
• sound_id 86161. [263168, 1253213] vs [173056, 1249146].
  • file name: Bus(Drive_Reverse)_1-2.wav

Between valid and dev sets:

• sound_id 86163. Its whole segment was used in both.
  • file name: City Ambience w_ Car Passing_1-2.wav
• sound_id 130603. [0, 822465] vs [11264, 768067].
  • Their file names vary. (Greek Chat2 - (Apollonia__39_s sPA) 18_44 05.10.wav // Greek Chat2 - (Apollonia's sPA) 18_44 05.10.wav)

Between eval and dev sets:

• soind_id 137692. [635904, 1430834] vs [141824, 1370009].
  • FREEZER_DOOR_OPEN_CLOSE.wav

I'm now wondering if this is a known issue to the (DCASE?) community / researchers. If so, do you know what is the recommended way to handle this? Otherwise, perhaps it's something that can be updated in Clotho 2.2 or something.

Hi! According to the description of the dataset there are 4981 audio samples in development split, but actually there are only 2893. Why?