Official implementation of

• Heavily Augmented Sound Event Detection utilizing Weak Predictions (DCASE2021 Challenge Task 4 technical report)
  by Hyeonuk Nam, Byeong-Yun Ko, Gyeong-Tae Lee, Seong-Hu Kim, Won-Ho Jung, Sang-Min Choi, Yong-Hwa Park
  
  - arXiv version has updates on some minor errors
• FilterAugment: An Acoustic Environmental Data Augmentation Method (Accepted to ICASSP 2022)
  by Hyeonuk Nam, Seong-Hu Kim, Yong-Hwa Park
  
  

Ranked on [3rd place] in IEEE DCASE 2021 Task 4, and accepted to ICASSP 2022.c

Also, refer to Frequency Dynamic Convolution SED which is upgraded SED model from this work!

Filter Augment is an audio data augmentation method newly proposed on the above papers for training acoustic models in audio/speech tasks. It applies random weights on randomly selected frequency bands. For more details, refer to the papers mentioned above.

• This example shows two types of FilterAugment applied on log mel spectrogram of a 10-second audio clip. (a) shows original log mel spectrogram, (b) shows log mel spectrogram applied by step type FilterAugment (c) shows log mel spectrogram applied by linear type Filter Augment.
• Applied filters are shown below. Filter (d) is applied on (a) to result in (b), and filter (e) is applied on (a) to result in (c)

• Step type FilterAugment shows several frequency bands that are uniformly increased or decreased in amplitude, while linear type FilterAugment shows continous filter that shows certain peaks and dips.
• On our participation on DCASE2021 challenge task 4, we used prototype FilterAugment which is step type FilterAugment without hyperparameter minimum bandwith. The code for this prototype is defiend as "filt_aug_prototype" at utils/data_aug.py @ line 52
• Code for updated FilterAugment including step and linear type for ICASSP submission is defiend as "filt_aug" at utils/data_aug.py @ line 7

Python version of 3.7.10 is used with following libraries

• pytorch==1.8.0
• pytorch-lightning==1.2.4
• pytorchaudio==0.8.0
• scipy==1.4.1
• pandas==1.1.3
• numpy==1.19.2

other requrements in requirements.txt

You can download datasets by reffering to DCASE 2021 Task 4 description page or DCASE 2021 Task 4 baseline. Then, set the dataset directories in config yaml files accordingly. You need DESED real datasets (weak/unlabeled in domain/validation/public eval) and DESED synthetic datasets (train/validation).

You can train and save model in exps folder by running:

python main.py

There are 5 configuration files in this repo. Default setting is optimized linear FilterAugment with ICASSP setting described in paper submitted to ICASSP. There are 4 other model settings from DCASE tech report with prototype FilterAugment. To train for model 1, 2, 3 or 4 from the DCASE tech report, you can run the following code instead.

# for example, to train model 3:
python main.py --confing dcase_model3

• These results are based on train models with single run for each setting

• These results are based on max values of each metric for 3 separate runs on each setting (refer to paper for details).

DCASE 2021 Task 4 baseline

If this repository helped your works, please cite papers below!

@techreport{nam2021heavilyaugmnetedsed,
    Author = "Nam, Hyeonuk and Ko, Byeong-Yun and Lee, Gyeong-Tae and Kim, Seong-Hu and Jung, Won-Ho and Choi, Sang-Min and Park, Yong-Hwa",
    title = "Heavily Augmented Sound Event Detection utilizing Weak Predictions",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
}

@INPROCEEDINGS{nam2021filteraugment,
    author={Nam, Hyeonuk and Kim, Seong-Hu and Park, Yong-Hwa},
    booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
    title={Filteraugment: An Acoustic Environmental Data Augmentation Method}, 
    year={2022},
    pages={4308-4312},
    doi={10.1109/ICASSP43922.2022.9747680}
}

Please contact Hyeonuk Nam at frednam@kaist.ac.kr for any query.