agkphysics commited on
Commit
196c090
1 Parent(s): e1a0de8

Update README.

Browse files
Files changed (1) hide show
  1. README.md +29 -8
README.md CHANGED
@@ -26,16 +26,37 @@ dataset_info:
26
  ---
27
 
28
  # AudioSet data
29
- This repository contains the balanced training set and evaluation set
30
- of the [AudioSet data](
31
  https://research.google.com/audioset/dataset/index.html). The YouTube
32
- videos were downloaded in March 2023, and so not all of the original
33
- audios are available.
34
 
35
- Extracting the `*.tar` files will place audio clips into the `audio/`
36
- directory. The distribuion of audio clips is as follows:
37
- - `audio/bal_train`: 18685 audio clips out of 22160 originally.
38
- - `audio/eval`: 17142 audio clips out of 20371 originally.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
39
 
40
  Most audio is sampled at 48 kHz 24 bit, but about 10% is sampled at
41
  44.1 kHz 24 bit. Audio files are stored in the FLAC format.
 
26
  ---
27
 
28
  # AudioSet data
29
+ This repository contains the balanced training set and evaluation set of
30
+ the [AudioSet data](
31
  https://research.google.com/audioset/dataset/index.html). The YouTube
32
+ videos were downloaded in March 2023, so not all of the original audios
33
+ are available.
34
 
35
+ The distribuion of audio clips is as follows. In parentheses is the dict
36
+ key used for HugginFace `datasets`:
37
+ - `bal_train` (`train`): 18685 audio clips out of 22160 originally.
38
+ - `eval` (`test`): 17142 audio clips out of 20371 originally.
39
+
40
+ You can use the `datasets` library to load this dataset, in which case
41
+ the raw audio will be returned along with a sequence of one or more
42
+ labels. Note that the raw audio is returned without further processing,
43
+ so you will need to decode and possibly downsample the audio for model
44
+ training.
45
+
46
+ Example instance from the `train` subset:
47
+ ```python
48
+ {
49
+ 'video_id': '--PJHxphWEs',
50
+ 'audio': {
51
+ 'path': 'audio/bal_train/--PJHxphWEs.flac',
52
+ 'array': array([-0.04364824, -0.05268681, -0.0568949 , ..., 0.11446512,
53
+ 0.14912748, 0.13409865]),
54
+ 'sampling_rate': 48000
55
+ },
56
+ 'labels': ['/m/09x0r', '/t/dd00088'],
57
+ 'human_labels': ['Speech', 'Gush']
58
+ }
59
+ ```
60
 
61
  Most audio is sampled at 48 kHz 24 bit, but about 10% is sampled at
62
  44.1 kHz 24 bit. Audio files are stored in the FLAC format.