Merge pull request #84 from YuchenHui22314/main

keyword_to_caption_debias
LAION-AI · Jan 20, 2023 · 572ebd2 · 572ebd2
2 parents ffd7f74 + 8a8ce47
commit 572ebd2
Showing 1 changed file with 6 additions and 1 deletion.
diff --git a/laion-audio-630k/README.md b/laion-audio-630k/README.md
@@ -30,7 +30,7 @@ We provide two version of Freesound dataset.
 |--------------------------------------------------|-------------------------|--------------------|---------                     |--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------- |
 | Freesound (no overlap)                           |2817.31hrs               | 460801             |1-2 captions per audio, audio    | [website](https://freesound.org/) <br> [**licenses file**](https://drive.google.com/file/d/1xF3K5x0RAhBNGKSMvE13cuvrIZLs6M3K/view?usp=share_link) <br> [csv_train+test](https://drive.google.com/file/d/1sm9pjPBEdXe1qGaGkiPRPj0Dq9cv0DPR/view?usp=sharing)<br>[csv_train](https://drive.google.com/file/d/1xekQ_mR_8-qEyzXmn5G7CGnWHb36XKdH/view?usp=sharing)<br>[csv_test](https://drive.google.com/file/d/1k7CnYjbkFZxNhtiP0vA7zLbM72xuG47G/view?usp=sharing)|[data card](/data_card/freesound.md)|
 | Freesound (full)                                 |3033.38hrs               | 515581             |1-2 captions per audio, audio    | [website](https://freesound.org/) <br> [**licenses file**](https://drive.google.com/file/d/1xF3K5x0RAhBNGKSMvE13cuvrIZLs6M3K/view?usp=share_link)<br> [csv](https://drive.google.com/file/d/10LRzpJN7CweCceuI_rXKpUafzilGFAir/view?usp=sharing)                                                |[data card](/data_card/freesound.md)|
-| Epidemic Sound                                   |220.41hrs                | 75645              |2 captions per audio, audio    | [website](https://www.epidemicsound.com/sound-effects/) <br> [csv](https://drive.google.com/file/d/1og3gk2V1t52XSPStpJECJ4OzfDMFX3Do/view?usp=sharing)                                |[data card](/data_card/Epidemic_sound.md)           |     
+| Epidemic Sound                                   |220.41hrs                | 75645              |2 captions per audio, audio    | [website](https://www.epidemicsound.com/sound-effects/) <br> [csv (Including T5-generated de-biased captions)](https://drive.google.com/file/d/1vo0NslkCTJHI03FbBSHLRztP6v2XkYNW/view?usp=sharing)                                |[data card](/data_card/Epidemic_sound.md)           |     
 | Audiostock                                       |46.30hrs                 | 10000              |1 caption per audio, audio    | [website](https://audiostock.net/se) <br> [csv](https://drive.google.com/file/d/1FnOcrb6fREIDBzB2lknJnszVn-yNCPp6/view?usp=sharing)                                                           |[data card](/data_card/Audiostock.md) |                
 | BBC Sound Effects                                |463.48hrs                | 15973              |1 caption per audio, audio    | [website](https://sound-effects.bbcrewind.co.uk/) <br> [csv*(no longer available, click to see explication below)](#about-bbc-sound-effects)                                                                                                                                |[data card](/data_card/BBC.md)| 
 | Free To Use Sounds                               |175.73hrs                | 6370               |Filename as caption, audio    | [website(need purchasing)](https://www.freetousesounds.com/product/all-in-one-sound-library-bundle/)                                                                                          |                                  | 
@@ -41,7 +41,12 @@ We provide two version of Freesound dataset.
 #### *About BBC Sound Effects
 Recently, BBC sound effects have modified their website structure. In consequence, only 300 samples are available for download. So, unfortunately, we are no longer able to generate csv file using our old scripts. In the meantime, many scrappers exist on GitHub, such as https://github.com/alisomay/bbc-sound-effects-downloader. You may try them to see if they work.
 
+## Keyword-to-Caption Augmentation
+We employ the keyword-to-caption model to augment labels of AudioSet and Epidemic Sound into corresponding captions with aid of a pre-trained language model T5. We also de-bias these captions by replacing, for example, "woman" and "man" with "person", aiming to eliminate potential gender discrimination. We hereby release the augmented captions for Epidemic Sound and AudioSet (in csv format).
 
+| Epidemic Sound | AudioSet |
+|----------------|----------|
+|[Epidemic_all_debiased.csv](https://drive.google.com/file/d/1vo0NslkCTJHI03FbBSHLRztP6v2XkYNW/view?usp=sharing) | [csv files for AudioSet balanced_train, unbalanced_train, and eval splits](https://drive.google.com/drive/folders/1Ip5Eh_Mb7Ye09ZpqKA-dTIdh4SgFDSnB?usp=sharing) |
 ## Credits & Licence
 - **!!!TERM OF USE!!!**: **By downloading audios through the links provided in the csv files, you agree that you will use the audios for research purposes only, unless you get the permission from owners of the Datasource that you can use it for other purposes.**
 ### Freesound Credit: