Utility functions for preprocessing PodcastFillers dataset and code for reproducing Table1 and Table2 in the FillerNet paper with AVC-FillerNet sed_eval predictions.

Dataset homepage: PodcastFillers
Dataset zenodo page: Zenodo

tqdm==4.61.2
sed_eval==0.2.1
dcase_util==0.2.18
pandas==1.1.5

In preprocessing script, we first convert full-length MP3 podcast episodes into WAVs, then we cut 1-second event clips based on the meta csv with converted WAVs. Format conversion:

python preprocessing_script.py -dataset_path {dataset_path} -stage reformat

Event clip WAV cut:

python preprocessing_script.py -dataset_path {dataset_path} -stage cut

We prepare two customized parameters to preprocessing event clips:

• SAMPLE_RATE : Sampling rate for the converted WAV files, default value is 16kHz;
• DURATION: Length of the event clips(unit: second), the filler/non-filler event will also be centered in the clip, the default value is 1.0 and it is larger than zero.

To reproduce the AVCFIllerNet results from Table.1 and Table.2 from our paper, run

python reproduce_results.py -dataset_path {dataset_path}

reformat and generate_clip_wav passed pytest using pytest -q test.py.