This project objective is to evaluate models on a ABX task on noise.

We first evaluate the models on the Dcase Dataset. We use the STARSS22 dataset to generate the .item files.

The Sony-TAu Realistic Spatial Soundscapes 2022 (STARSS22) dataset contains multichannel recordings of sound scenes in various rooms and environments, together with temporal and spatial annotations of prominent events belonging to a set of target classes.

The 13 sound classes are :

0. Female speech, woman speaking
1. Male speech, man speaking
2. Clapping
3. Telephone
4. Laughter
5. Domestic sounds
6. Walk, footsteps
7. Door, open or close
8. Music
9. Musical instrument
10. Water tap, faucet
11. Bell
12. Knock

For each recording, the labels are provided in a CSV file : [frame number (int)], [active class index (int)], [source number index (int)], [azimuth (int)], [elevation (int)]

A frame correspond to a temporal relolution of 100ms.

We removed the "Music" class (8), and keep the frames with only one class activated at a time to compute the ABX scores.

The results on Dcase only are promising, so we decided to keep only a few classes and to extend the dataset with noise segments from AudioSet.

The classes kept from Dcase are :

• Walk, footsteps
• Clapping
• Water tap
• Male speech
• Female speech
• Domestic sounds : seperated in two classes Vacuum cleaner and Air conditioning

To separate domestic sounds, we used two additional labels on Dcase: 13. Vacuum Cleaner 14. Air Conditioning

We added the following classes from AudioSet:

• Air conditioning
• Baby Cry
• Knock
• Purr
• Rain
• Vacuum cleaner
• Walk, footsteps
• Water tap

The final item files are in : item_files/item_files_merged

To compute the ABX score, use CPC2.

You can use the launchers in launchers

The item files are in ./item_files/final_merged

The audiofiles are on Jean Zay : /gpfswork/rech/xdz/commun/abx_noise/audiofiles

You can use scripts/plot_abx.py to plot the mean and std ABX error rate according to training duration.