search

tensorflow / datasets

TFDS is a collection of datasets ready to use with TensorFlow, Jax, ...
https://www.tensorflow.org/datasets
Apache License 2.0
4.29k stars 1.54k forks source link

[data request] Google Research AudioSet #2399

Open whatwilliam opened 4 years ago

whatwilliam commented 4 years ago

This dataset contains 632 classes of audio events derived from 10-second clips from Youtube that have been human-labeled. The events include recognizing music, speech, vehicle(s), creaking, key-jangling, and so on. The dataset consists of the youtube link, its 10-second segment, and the one or more labels that are contained within that time range. The audio extracted using the youtube-dl library will be .mp3. The sound clips are also rated in terms of quality.

The dataset can be used to train models to predict and identify multiple sounds when presented with a sound clip. Because the dataset has a large number of classes, any sound clip input will likely have at least one label that applies to it. It also contains quality estimates that rate the sound quality of the clip for each class.

Folks who would also like to see this dataset in tensorflow/datasets, please thumbs-up so the developers can know which requests to prioritize.

And if you'd like to contribute the dataset (thank you!), see our guide to adding a dataset.

infiniteoverflow commented 4 years ago

I would like to contribute this dataset. I have personally worked using this dataset for my personal project. Can you please assign me this issue? Thanks !!

whatwilliam commented 4 years ago

I would like to contribute this dataset. I have personally worked using this dataset for my personal project. Can you please assign me this issue? Thanks !!

Hey @infiniteoverflow, sorry I didn't see your post. I've been working on Audioset as well and I think our PRs are synergistic Something I wanted was the raw audioclip data from youtube, but this was difficult to get and required me to manually download youtube videos and strip to the 10s clip of interest.

Would you be interested in merging our versions together as a builder config?

Austronesia commented 2 years ago

Hello. I don't know how to get the "balanced train videos for singing" links. I only manage to find the page with the list. How should I get it ?

whatwilliam commented 2 years ago

Hello. I don't know how to get the "balanced train videos for singing" links. I only manage to find the page with the list. How should I get it ?

https://research.google.com/audioset/download.html You can download the related dataset in their 128-dimension format here or take the csv files and find all the examples with 'balanced train videos for singing labels' at this download link. http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/class_labels_indices.csv

Alternatively, you could take the csv of all the youtube IDs and use youtube-dl to take the video's audio data that you need.