This is great. It is really nice to see someone take the time and effort to create large usable dataset. Only the break through research and next big thing in neural nets seems to get all the attention. But an average Joe as no easy way of getting access to data sets.
I also like that they even leveraged stuff like "x for y hours" style videos on youtube. (Eg: Dog barking to 10 hours).
I remember Andrew Ng mention something about being able synthesize even more trainable data by using combination of data as noise. (Eg: Women talking + dog barking (as introduced noise)).
Good job google, keep it up. I wonder when I will see other major players publish large data sets? Maybe Microsoft is next?
Microsoft [1][2] and Google [0]have been doing this for years. Google's billion word language model dataset is a notable example which was central to a lot of important research.
"The ontology is made available by Google Inc. under a Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license." -- Github page[0]
Looks like the... tag names? and example urls have been released, but the videos and sound are under their respective licenses -- ie, mostly the standard Youtube license.
This is neat. Can a ML model developed on this dataset be used for commercial purposes? I guess, at minimum, the paper and tag list are provided as help for those corporations that would wish to build/use a private dataset for similar purposes?
I spent a few minutes browsing through the data, and immediately founds lots of mislabeled data, e.g. an industrial power washer labeled as a toothbrush sound.
7 comments
[ 4.5 ms ] story [ 26.6 ms ] threadI also like that they even leveraged stuff like "x for y hours" style videos on youtube. (Eg: Dog barking to 10 hours).
I remember Andrew Ng mention something about being able synthesize even more trainable data by using combination of data as noise. (Eg: Women talking + dog barking (as introduced noise)).
Good job google, keep it up. I wonder when I will see other major players publish large data sets? Maybe Microsoft is next?
[0] https://research.google.com/pubs/pub41880.html [1] https://www.microsoft.com/en-us/research/project/mslr/ [2] https://www.microsoft.com/en-us/research/project/microsoft-3...
Looks like the... tag names? and example urls have been released, but the videos and sound are under their respective licenses -- ie, mostly the standard Youtube license.
This is neat. Can a ML model developed on this dataset be used for commercial purposes? I guess, at minimum, the paper and tag list are provided as help for those corporations that would wish to build/use a private dataset for similar purposes?
[0] https://github.com/audioset/ontology
https://research.google.com/audioset/dataset/index.html