Self-training network for sound event detection

This is the implementation of self-training approach described in a paper "Self-Training for sound event detection in audio mixtures" published in ICASSP2021. This approach was implemented based on a baseline for the task4 Sound Event Detection (SED) in DCASE challenge 2020. Original implementation can be found in https://github.com/turpaultn/dcase20_task4.

• The next version of self-training is available in https://github.com/JHU-LCAP/CRSTmodel

Dependencies

Python >= 3.6, pytorch >= 1.0, cudatoolkit>=9.0, pandas >= 0.24.1, scipy >= 1.2.1, pysoundfile >= 0.10.2, scaper >= 1.3.5, librosa >= 0.6.3, youtube-dl >= 2019.4.30, tqdm >= 4.31.1, ffmpeg >= 4.1, dcase_util >= 0.2.5, sed-eval >= 0.2.1, psds-eval >= 0.1.0, desed >= 1.1.7

Dataset

In order to train a network, the challenge dataset for SED task was used. The challenge dataset is categorized into three groups: strong labeled data, weakly labeled data, and unlabeled data and they are available in https://project.inria.fr/desed/.

• training data

1. weakly labeled data (DESED, real:weakly labeled): 1,578
2. unlabeled data (DESED, real:unlabeled): 14,412
3. strong labeled data (DESED, synthetic:training): 2,595

• test data

1. validation data (DESED, real:validation): 1,168

Usage

Same with the challenge baseline.

Step 1. download the dataase

Step 2. modify data path in "config.py"

Step 3. run "main_plabel.py"

Citation

S. Park, A. Bellur, D. K. Han, and M. Elhilali, "Self-training for sound event detection in audio mixtures," in proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2021, pp. 341-345.