The scripts defines the different steps to download the data.
-
1- It downloads the data (real data and synthetic soundbank):
- Download recorded data corresponding to DESED dataset. (see contacts in repo README for missing_files)
- Soundbank training from DESED, see DESED_synth
- Download background training to be included in soundbank
-
2- Generate the base synthetic data:
- Download the Jams from 2020 on the DESED dataset and create audio files (using Scaper).
-
3- (NOT USED FOR THE BASELINE) Reverberate synthetic data:
- Download RIR from FUSS
- Reverberate the synthetic data generated in 2)
- Reverberated data are used to train the baseline not using Source separation.
-
4- Separate sounds using FUSS:
- Download baseline model from Fuss
- Apply the FUSS baseline model on the synthetic soundscapes + recorded soundscapes from DESED. These data are used to train the baseline using Source separation.
generate_new synthetic_data.sh is the example used to generate the synthetic data.
Feel free to take example on this and check generate_synth_dcase20.py to modify the parameters.