ConformerSED

Implementation of the paper CONFORMER-BASED SOUND EVENT DETECTION WITH SEMI-SUPERVISED LEARNING AND DATA AUGMENTATION

Dataset

To run this repository's script, please make sure the dataset is cofigured as follows:

└── dcase20_task4
     └── dataset
         └── audio
             ├── train
             |   ├── synthetic20
             |   ├── weak
             |   └── unlabel_in_domain
             ├── validation
             └── public

To acquire the dataset, run stage 1 in bin/run.sh or download directly.

Quick start

$ conda env create -f environment.yaml
$ conda activate dcase20t4
$ wandb login <YOUR_API_KEY>
$ bash ./bin/run.sh

Results

	Event-based macro F1	PSDS
Validation	47.7%	0.637
Public	49.0%	0.681

To reproduce the result on public eval set, download pretrained model and config file here.

Then, put model and config as follows:

├── data
|    └── feat
|        └── sr16000_n_mels64_n_fft1024_n_shift323
|            └── public
└── exp
    └── conformer_sed                                  
        ├── model
        │   └── model_best_score.pth
        ├── model_config.yaml
        ├── trainer_config.yaml
        ├── post_process_params.pickle
        └── stats.npz

After that, you can run test script

$ python ./src/test.py

Disclaimer

• Some audio samples in the original dataset may be missing for some reason (e.g., removed on the internet). For audio samples used in this experiment, see metadata.
• This repository results are slightly better than reported in the paper.

Contact

If you have any questions, please feel free to ask me.

Koichi Miyazaki (E-mail: miyazaki.koichi_at_g.sp.m.is.nagoya-u.ac.jp)

Reference

DCASE2020 Task4 baseline: https://github.com/turpaultn/dcase20_task4