audioset / ontology Goto Github PK

Bassoon's (/m/01c3q) citation_uri: http://en.wikipedia.org/wiki/Oboe is not correct. It should be: https://en.wikipedia.org/wiki/Bassoon.

I propose this amendment on behalf of the Freesound Datasets team.

"Buzz" (/m/07pjwq1) parents are:

1. Animal > Wild animals > Insect > Fly, housefly
2. Animal > Wild animals > Insect > Bee, wasp, etc.
3. Source-ambiguous sounds > Onomatopoeia > Brief tone

And its definition is: "The sound of rapid vibration, commonly the wings of a flying insect."

Therefore the "Buzz" class does not seem to include, e.g., "Engine" (/m/02mk9) or "Guitar amplifier" (/m/01vfsf). Is it worth including these parents? Or these kinds of sounds should not be included into "Buzz" (/m/07pjwq1)?

In case of willing to include more kinds of sounds into "Buzz", we think that it is more correct that it is an "Onomatopoeia" (/m/05n1m) and, as such, it is a sibling of "Hum" (/m/07rcgpl) - which is very similar. In this case, shall we change the parent "Source-ambiguous sounds > Onomatopoeia > Brief tone" to a more generic parent "Source-ambiguous sounds > Onomatopoeia"?

I propose this discussion on behalf of the Freesound Datasets team.

I know that the ontology is meant to evolve over time, but for archival purposes, it would be useful to have tagged revisions (ie, v1) and DOI's (via Zenodo).

Is there any chance of this happening?

For example: dragons, demons, banshees, aliens, things like that? This is an issue for sound effects libraries.

Hi everyone

The following entities of the temporally strong labeled part of Audioset are missing:

0. /m/0174k2 Washing machine
1. /m/018p4k Cart
2. /m/01j2bj Bathroom sounds
3. /m/01j3j8 Studio recording, Music
4. /m/01lynh Stairs
5. /m/02417f Windscreen wiper, windshield wiper
6. /m/0269r2s Chain
7. /m/02f9f_ Shower
8. /m/02ll1_ Lock
9. /m/040b_t Refrigerator
10. /m/04ctx Knife
11. /m/056r_1 Keypress tone
12. /m/0641k Paper rustling
13. /m/06cyt0 Mechanical bell
14. /m/07pqmly Slurp, drinking straw
15. /m/07s13rg Sweeping
16. /m/07sk0jz Stomp, stamp
17. /m/08dckq Carbon monoxide detector, CO detector
18. /m/098_xr Error signal
19. /m/0bcdqg Ringing tone, ringback tone
20. /m/0bzvm2 Video game sound
21. /m/0c1tlg Electric rotor drone, quadcopter
22. /m/0d4wf Kitchen and dining room sounds
23. /m/0fw86 Tap dance
24. /m/0hgq8df Crockery breaking and smashing
25. /m/0md09 Power saw, circular saw, table saw
26. /t/dd00138 Brief tone
27. /t/dd00141 Pant (dog)
28. /t/dd00142 Audio logo
29. /t/dd00143 Unknown sound
30. /t/dd00144 Alert
31. /t/dd00147 Dong, bong

Do you have plans to add them to the ontology?

Hi Dan:
Sorry for bugging again!
"The file audioset_eval_strong.tsv describes 139,538 segments across the 16,996 excerpts from the evaluation set. There are 416 MIDs, 9 of which are not present in the train labels."

In my experiment, there seem to be 35 MIDS that are different than the original weakly labeled 527 MIDS:
{'/m/0bzvm2', '/t/dd00139', '/t/dd00098', '/m/07q8f3b', '/m/0c1tlg', '/m/0md09', '/t/dd00091', '/m/093_4n', '/m/01sb50', '/m/0174k2', '/m/01j423', '/m/0hgq8df', '/t/dd00099', '/m/05mxj0q', '/t/dd00141', '/m/01lynh', '/m/0fw86', '/m/0dgw9r', '/t/dd00061', '/t/dd00109', '/m/09l8g', '/m/07sk0jz', '/t/dd00133', '/m/0d4wf', '/m/018p4k', '/t/dd00143', '/m/0bcdqg', '/m/09hlz4', '/m/0zmy2j9', '/t/dd00138', '/t/dd00142', '/m/02f9f_', '/m/02021', '/m/01j3j8', '/m/0641k'}

Just want to confirm this is the expected behavior.

This is very similar to Issue 6, but not quite the same.

Thank you in advance!

Hi, there:
The Quality Assessment and rerating file at:
https://research.google.com/audioset/download.html
is not downloadable.
Namely, qa_true_counts.csv rerated_video_ids.txt are both not available. Could you fix this?

Thank you!

Hi, there:

Just wondering how to decode the labels:

/m/06mb1,/m/0jb2l,/m/0ngt1,/t/dd00038

Thanks.

Hi, Dan and the other contributors:
Thank you for maintaining the repo so far!
AudioSet, to me, is a great resource, and still the best resource to understand the nature of sound.
We did a recent study: paper and code, where we found the recent research papers are having a whooping +- 5% difference in performance due to test set missing files when downloading. Plus, the difference in label quality also contributed to the performance variation and making it less fair (see figure 2 in our paper).
I understand you guys have legal constraints on youtube licensing, but guess this issue could be easier for the original authors to address. Either to advocate the community to use a common subset, or release an updated test set? given you guys already released updated strong labels.
Looking forward to your thoughts.