Question about ID mapping between AudioSkills and original AudioSet

[zhuangweiji]

Hi, thanks for releasing the great nvidia/AudioSkills dataset!

I’m currently exploring the dataset (e.g., audioskills_xl/AudioSet.json) and noticed that the "id" fields (e.g., "YJRaxh5RfawI", "YFHuxuM-iRo4") don’t seem to correspond to any YouTube IDs in the official AudioSet CSVs (balanced_train_segments.csv, unbalanced_train_segments.csv, eval_segments.csv) from AudioSet.

I checked using:

grep -E "YJRaxh5RfawI|YFHuxuM-iRo4" Audioset/{eval_segments.csv,unbalanced_train_segments.csv,balanced_train_segments.csv}

No matches were found.

My questions:
• Are these IDs derived from original AudioSet YouTube IDs?
• If not, what’s the correct way to align AudioSkills samples with entries from the original AudioSet (for example, to obtain labels or categories)?
• Is there a mapping file between AudioSkills IDs and original AudioSet segment IDs?

Any clarification would be greatly appreciated — thanks again for this excellent work!

[zhuangweiji]

I also noticed that some entries in AudioSet.json appear to contain duplicate QA pairs for the same audio file and prompt. Here’s an example (see image below):

[zhuangweiji]

Okay, it seems that the "id" field in AudioSet.json might be the original YouTube ID with a Y prefix (e.g., "YJRaxh5RfawI" → "JRaxh5RfawI").
But why are there duplicate entries in AudioSet.json?

[Sreyan88]

You are correct about the AudioSet mapping, where you need to remove the leading "Y" to get the mapping. This is common nomenclature among many datasets. Thank You for the comment, will add a note!

I need to check about the duplicate entries. Maybe you can ignore them for now. Do you know how many repeated entries there are? If only a few, maybe I can just put a note. If significant, will take a look.

[zhuangweiji]

Hi,

Thanks for the clarification! I checked the AudioSet.json file — there are 467,725 completely duplicated entries (i.e., all fields are exactly the same), out of a total of 2,790,508 entries.

So roughly 16.8% of the entries are exact duplicates.